ROBUST STATOR INSULATION PROGNOSIS TECHNIQUE FOR

INVERTER-DRIVEN MACHINES

By

William Robert Jensen

A DISSERTATION

Submitted to

Michigan State University

in partial fulﬁllment of the requirements

for the degree of

Electrical Engineering - Doctor of Philosophy

2019

ABSTRACT

ROBUST STATOR INSULATION PROGNOSIS TECHNIQUE FOR

INVERTER-DRIVEN MACHINES

By

William Robert Jensen

Stator insulation degradation is one of the most common causes for failure in an electric

machine. As wide bandgap device become more popular for inverter topologies, electrical

stress increases and insulation failure becomes a more signiﬁcant concern. Short circuits

formed from degraded insulation can quickly lead to a catastrophic failure. A technique

to detect when insulation is degraded, well before the formation of a short circuit, allows

the machine to be safely powered-down. In this work, an online technique to detect insula-

tion degradation and provide a failure prognosis is proposed. The proposed technique does

not require high-frequency sampling or additional sensors as these requirements are costly.

Accelerated thermal degradation of stator insulation is performed experimentally and the

results show a trend in the measured current that can be used for prognosis.

Inverter switching devices also degrade over their lifetime. A switching device that ex-

periences gate oxide degradation produces features in the measured current that can mask

changes due to insulation degradation. In this work, an online technique to detect gate oxide

degradation in inverter switching devices is proposed. Accelerated gate oxide degradation of

silicon and silicon-carbide MOSFETs shows that there are two features in the current that

appear as the device degrades. Experimental results verify that one of the features in the

current can be detected using steady-state voltage commands. Detecting degradation using

quantities that are already calculated in the controller eliminates the need for additional

sensor or high-frequency sampling.

An algorithm to improve the robustness of the insulation failure prognosis is proposed.

As gate oxide degradation can mask insulation degradation, it can also lead to an under-

estimation of remaining useful life. Also, as there are many sources of stress that degrade

insulation, a change in the rate of degradation due to diﬀerences in applied stress or in the

insulation system can signiﬁcantly impact the insulation lifetime. The proposed technique

improves robustness of the insulation failure prognosis by ﬁrst separating between insulation

and switching device degradation. Once it is determined that the insulation is degrading, the

remaining useful life is predicted using the proposed algorithm that is robust to varying rates

of degradation as well as variations in the insulation system. Data sets from experimental

insulation degradation are used to compare the accuracy and robustness of the proposed

stator insulation prognosis algorithm.

To my wife Gabriela and parents Jan and Steve

iv

ACKNOWLEDGMENTS

I would like to express my sincere appreciation to Dr. Shanelle Foster, whose constant

guidance, advice and motivation made it possible for me to complete my PhD and helped

me develop as an engineer and researcher. I also want to wish a special thanks Dr. Elias

Strangas, Dr. Bingsen Wang and Dr. Guoming Zhu for their insights and guidance to

improve the quality of this dissertation.

I would like to thank to my colleagues from the Electrical Machines and Drives Lab-

oratory, Cristi´an L´opez-Mart´ınez, Dr. Reemon Haddad, Dr. Andrew Babel, Dr. Rodney

Singleton, Anmol Aggarwal, Shaopo Huang, Orwell Madovi, Tiraruek Ruekamnuaychok,

Steve Hayslett, Matt Meier, Shubham Shedge, Prathima Nuli, Ibrahim Allaﬁ, Tia Smith,

Josh Ward and Heinrich Eickhoﬀ. This work would not be possible without their advice

and assistance.

I would like to oﬀer a special thanks to Thang Pham, whose generous

assistance was instrumental for me to complete this work. I also want acknowledge Dr. No-

mar Gonz´ales-Santini, Dr. Ameer Janabi, Dr. Allan Taylor and Dr. Ujjwal Karki for the

assistance they provided.

I would also like to thank Brian Wright, Meagan Kroll, Laurie Rashid, Roxanne Peacock,

Gregg Mulder and Michelle Stewart from the ECE department at Michigan State University

for all of their assistance.

Additionally, I want to thank the James Dyson Foundation for their ﬁnancial support

during my study.

I want to express my deep gratitude to my parents Jan and Steve, siblings Ben and

Bonnie for their strong support and assistance for me in preparing this work. I also want

to express my appreciation to my father-in-law Jes´us and mother-in-law Elo´ısa, as well as

v

Jes´us, Patricia, Anamar´ıa and Jorge, for their encouragement and support.

This work, and all of my studies, would not have been possible without the love, patience

and much-needed support from my wife Gabriela. Her encouragement was crucial in com-

pletion of this work. I also want to thank Enzo, for providing comfort as I completed this

work.

vi

TABLE OF CONTENTS

LIST OF TABLES . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

ix

LIST OF FIGURES . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xii

KEY TO ABBREVIATIONS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xviii

Chapter 1

Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.1 Objectives and contributions . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.2 Organization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

Chapter 2

Insulation Degradation . . . . . . . . . . . . . . . . . . . . . . . . .
2.1
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.2 Simulation results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

Insulation model

Chapter 3

Insulation Degradation Detection Technique . . . . . . . . . . .
3.1 Proposed technique . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.2 Experimental setup . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.2.1
Simulated degradation . . . . . . . . . . . . . . . . . . . . . . . . . .
3.2.2 Accelerated thermal degradation . . . . . . . . . . . . . . . . . . . .
3.2.3 Machines under test
. . . . . . . . . . . . . . . . . . . . . . . . . . .
3.2.4 Failure criteria . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.2.5 Data acquisition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

3.3 Experimental results

Chapter 4 Online Gate Oxide Degradation Detection . . . . . . . . . . . .
4.1 Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.2 Current Techniques . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.3 Eﬀects of gate oxide degradation . . . . . . . . . . . . . . . . . . . . . . . .
4.4 Accelerated degradation test . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.4.1 Test bed . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.4.2 Eﬀects of gate oxide degradation . . . . . . . . . . . . . . . . . . . .
4.5 Proposed technique . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.6 Online detection experiment . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.6.1 Experimental setup . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.6.2 Results with simulated degradation . . . . . . . . . . . . . . . . . . .
4.7 Detection at higher switching frequencies . . . . . . . . . . . . . . . . . . . .

Chapter 5 Robust Insulation Failure Prognosis Algorithm . . . . . . . . .
5.1 Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
5.1.1 Extended Kalman Filter . . . . . . . . . . . . . . . . . . . . . . . . .

vii

1
2
4

5
9
14

17
20
22
22
23
26
28
29
32

36
36
38
40
42
43
46
52
55
58
60
62

65
65
67

5.2 Prognosis

5.1.2 Particle Filter . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
5.2.1 Metrics of accuracy . . . . . . . . . . . . . . . . . . . . . . . . . . . .
5.2.2 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
5.2.3 Computational eﬀort analysis
. . . . . . . . . . . . . . . . . . . . . .
5.2.4 Reduced sampling rate . . . . . . . . . . . . . . . . . . . . . . . . . .
5.3 Evaluation of robustness . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
5.3.1
Inaccurate training data . . . . . . . . . . . . . . . . . . . . . . . . .
5.3.2 Varying rate of degradation . . . . . . . . . . . . . . . . . . . . . . .
5.3.3 Diﬀerent insulation system . . . . . . . . . . . . . . . . . . . . . . . .
5.4 Failure threshold . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
5.5 Proposed prognosis algorithm . . . . . . . . . . . . . . . . . . . . . . . . . .

69
73
76
78
81
83
84
85
87
88
89
91

Chapter 6 Fault Separation . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
6.1 Proposed technique for fault separation . . . . . . . . . . . . . . . . . . . . .

96
96

Chapter 7 Conclusion and Future Work . . . . . . . . . . . . . . . . . . . . . 101

Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104

viii

LIST OF TABLES

Table 2.1

Thermal class designations for insulation systems . . . . . . . . . . .

6

Table 2.2

Parameters of the insulation machine modeled . . . . . . . . . . . .

15

Table 3.1

List of standard insulation tests with their drawbacks

. . . . . . . .

18

Table 3.2

Ratings of PMSM used to collect leakage current measurements . . .

22

Table 3.3

Operating point at which leakage current is measured . . . . . . . .

22

Table 3.4

Number of turns between diﬀerent labeled terminals . . . . . . . . .

23

Table 3.5

Component operating parameters used in analog peak detector circuit 26

Table 3.6

Ratings of the machines used in accelerated thermal degradation testing 27

Table 3.7

Visible geometric diﬀerences between the stators and insulation in the
two batches of motors used in accelerated degradation experiments
where Batch 2 motors are the higher eﬃciency designs . . . . . . . .

Table 3.8

Temperatures and duration applied for accelerated thermal degrada-
tion tests . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

Table 3.9

Sampling rate of oscilloscope to measure current across in-series re-
sistor and to measure output of peak detector circuit . . . . . . . . .

28

33

35

Table 4.1

Values of parameters used in simulation . . . . . . . . . . . . . . . .

43

Table 4.2

Properties of the IRF520 power MOSFET and C3M0280090D SiC
MOSFET . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

Table 4.3

Details of components and setup information for power Si MOSFET
accelerated degradation testing . . . . . . . . . . . . . . . . . . . . .

Table 4.4

Details of components and setup information for SiC MOSFET ac-
celerated degradation testing . . . . . . . . . . . . . . . . . . . . . .

Table 4.5

Diﬀerent load types and the features measured when the MOSFET
applied voltage across each . . . . . . . . . . . . . . . . . . . . . . .

44

46

46

46

ix

Table 4.6

Experimental changes in turn-on transient characteristics from Si
power MOSFET . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

50

Table 4.7

Table of SiC MOSFET results . . . . . . . . . . . . . . . . . . . . .

51

Table 4.8

Simulated operating point for comparing calculated duty cycles with
diﬀerent delay times . . . . . . . . . . . . . . . . . . . . . . . . . . .

57

Table 4.9

Experimental steady-state operating conditions . . . . . . . . . . . .

59

Table 4.10

Parameters of PMSM used for experimental validation . . . . . . . .

60

Table 4.11

Percent change in calculated duty cycle for each voltage vector with
100nsec and 500nsec delay added in high-side switch of phase A . .

Table 4.12

Percent change in calculated duty cycle for each voltage vector with
500nsec delay added in high-side switch of speciﬁed phase . . . . . .

Table 4.13

Simulated operating point for comparing calculated duty cycles with
diﬀerent delay times . . . . . . . . . . . . . . . . . . . . . . . . . . .

Table 4.14

Comparison of normalized change in duty cycles with diﬀerent delay
and switching frequency (fs) simulated . . . . . . . . . . . . . . . .

Table 5.1

Initial trend parameters and settings used in EKF and PF to estimate
RUL for Machines 2 and 3 . . . . . . . . . . . . . . . . . . . . . . .

61

62

63

64

79

Table 5.2

Accuracy of RUL estimated using EKF and PF for Machines 2 and 3

80

Table 5.3

Accuracy of last 25% of estimated RUL using EKF and PF for Ma-
chines 2 and 3 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

80

Table 5.4

Steps required to implement the EKF . . . . . . . . . . . . . . . . .

81

Table 5.5

Steps required in implementing the PF . . . . . . . . . . . . . . . .

81

Table 5.6

Required memory to implement the EKF . . . . . . . . . . . . . . .

82

Table 5.7

Required memory to implement the PF . . . . . . . . . . . . . . . .

82

Table 5.8

Comparison of EKF and PF computational cost with Ns = 2379 . .

82

x

Table 5.9

Comparison of the rate of decrease from the peak-to-peak trend, pos-
itive peak trend and analog circuit output trend for the last 70 hours
of life from Machine 4 . . . . . . . . . . . . . . . . . . . . . . . . . .

Table 5.10 Accuracy of RUL estimated using EKF and PF without using accu-
rate training data for Machines 2 and 3 . . . . . . . . . . . . . . . .

Table 5.11 Accuracy of last 25% of RUL estimated using EKF and PF without
accurate training data for Machines 2 and 3 . . . . . . . . . . . . . .

Table 5.12 Accuracy of RUL estimated using EKF and PF for Machines 1 over
the whole lifetime and the last 50% of life, where data sets from
Machines 2 and 3 are used for training . . . . . . . . . . . . . . . . .

Table 5.13 Accuracy of RUL estimated using EKF and PF for Machine 5 over the
whole lifetime and the last 60% of life, where data sets from Machines
2 and 3 are used for training . . . . . . . . . . . . . . . . . . . . . .

Table 5.14 Accuracy of RUL estimated using EKF, PF and proposed algorithm
for Machines 2 with and without accurate training data . . . . . . .

Table 5.15 Accuracy of RUL estimated using EKF, PF and proposed algorithm
for Machines 3 with and without accurate training data . . . . . . .

Table 5.16 Accuracy of RUL estimated using EKF, PF and proposed algorithm
for Machines 1 using data sets from Machines 2 and 3 as training . .

Table 5.17 Accuracy of RUL estimated using EKF, PF and proposed algorithm
for Machines 5 using data sets from Machines 2 and 3 as training . .

84

85

87

88

89

92

93

93

95

xi

LIST OF FIGURES

Figure 2.1

Eﬀect of thermal stress on dielectric strength [1]

. . . . . . . . . . .

Figure 2.2

Eﬀect of voltage stress on time to reach 5th of 10 failures [1]

. . . .

Figure 2.3

Simpliﬁed model of groundwall insulation between energized conduc-
tor in the stator slot and the grounded back iron . . . . . . . . . . .

Figure 2.4

As insulation degrades, physical changes lead to a decrease in overall
area of material around conductors
. . . . . . . . . . . . . . . . . .

9

9

10

11

Figure 2.5

Transmission line model of one phase winding implemented in Simulink,
where the capacitance and resistance to ground represents the insulation 11

Figure 2.6

Voltage measured from phase-to-ground at the terminal and near the
neutral point in a simulated inverter-driven, three-phase winding . .

Figure 2.7

One of the several transmission-line models, used to represent a phase
winding, used for analyzing the change in leakage current overshoot

Figure 2.8

Output shows varying overshoot magnitudes as ζ varies where input
to system is step response . . . . . . . . . . . . . . . . . . . . . . . .

11

12

14

Figure 2.9

One slot of random-wound induction machine with insulation included 15

Figure 2.10 Simulated voltage applied to healthy and degraded winding insulation
Insulation
. . . . . . . . . .

transmission line model and leakage current responses.
impedance is decreased by 50% in this simulation.

Figure 2.11 Model of one phase winding with turn-to-turn insulation modeled as
parallel resistance and capacitance . . . . . . . . . . . . . . . . . . .

Figure 2.12 Leakage current transient response shows 3.57% decrease at voltage
transient in simulation when turn-to-turn insulation impedance de-
creases by 50% . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

15

16

16

Figure 3.1 Winding schematic with 15nF capacitor inserted across 14 turns

. .

23

Figure 3.2

Environmental chamber used to degrade the insulation where electri-
cal connections can be made through the side of the chamber . . . .

25

xii

Figure 3.3

Experimental setup for measuring leakage current

. . . . . . . . . .

25

Figure 3.4

Experimental measurement setup with peak detector shown . . . . .

26

Figure 3.5

Healthy images of insulation in machines from two diﬀerent batches,
where the winding and amount of insulation between the two batches
are visually quite diﬀerent
. . . . . . . . . . . . . . . . . . . . . . .

Figure 3.6

Images of missing insulation in Machine 1, where the lighter area in
the winding is bare copper, after failure . . . . . . . . . . . . . . . .

Figure 3.7

Leakage current measurement on left, where zooming in and exam-
ining the leakage current at the voltage rise on the right shows the
transient oscillation and the extraction of the leakage current over-
shoot, Ileak . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

Figure 3.8

Raw signal and signal after noise is removed from experimental leak-
age current of machine with healthy insulation . . . . . . . . . . . .

Figure 3.9

Insulation current overshoot at diﬀerent levels of thermal degradation
where a decrease in the peak-to-peak value is measured . . . . . . .

Figure 3.10 Output signals from the leakage current measured across an in-series
. . . . . . . . . . . . . . . . .

resistor and the analog peak detector

Figure 3.11 Experimental measurement of leakage current with and without an
. . . . . . . . . .

additional 15nF capacitor inserted between turns

Figure 3.12 Trend in leakage current overshoot magnitude from healthy to failure
in all nine machines . . . . . . . . . . . . . . . . . . . . . . . . . . .

Figure 3.13 Trend in leakage current and peak detector output during testing of
Machine 4 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

27

29

30

30

31

31

32

34

35

Figure 4.1

Illustration of physical eﬀect of gate oxide degradation in MOSFET

37

Figure 4.2

Changes in turn-on transient voltages with gate oxide degradation .

38

Figure 4.3

Available current measurements in an inverter-driven machine. Leak-
age current to ground creates an imbalance in the three-phase currents. 40

Figure 4.4

Circuit used in simulation to quantify change in leakage current tran-
. . . . . . .
sient response with voltage applied at a variable dV/dt

41

xiii

Figure 4.5

Figure 4.6

Simulation result showing decrease in applied dV/dt to a motor phase
causes a decrease in peak-to-peak overshoot in the leakage current.
The green arrows indicate when the voltage reached its ﬁnal value,
showing the diﬀerence in rise time. . . . . . . . . . . . . . . . . . . .

Gate-to-source voltage in healthy (subscript 1) and degraded (sub-
script 2) condition. With a higher threshold voltage, the time when
the MOSFET is conducting, ton, occurs further from the initial gate
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
signal rise

Figure 4.7

Simulated circuit used to measure change in phase current rise as
. . . . . . . . . . . . . . . . .
MOSFET threshold voltage increases

Figure 4.8

Change in threshold voltage leading to a delay in current rise with
inductive load . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

41

42

42

43

Figure 4.9

HEF circuit to apply electrical stress to the gate of the MOSFET . .

44

Figure 4.10 Experimental circuit used to apply stress and measure transient char-
. . . . . . .

acteristics by switching connections on MOSFET leads

Figure 4.11 Example signals collected from the oscilloscope before and after noise
was removed . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

Figure 4.12 Gate-to-source voltage waveforms for healthy device and after several
periods of HEF testing . . . . . . . . . . . . . . . . . . . . . . . . .

45

47

47

Figure 4.13 Leakage current waveforms after several periods of HEF testing . . .

48

Figure 4.14 Drain current vs Vgs to measure Vth . . . . . . . . . . . . . . . . . .

48

Figure 4.15 Delay in current rise after several periods of HEF testing . . . . . .

49

Figure 4.16 Current rise and Vgs from healthy and degraded SiC MOSFET where
. . .

ﬁgure on right is zoomed-in view of green box area on left plot

Figure 4.17 LTSpice simulation using C3M0280090D model provided from Wolf-
speed where the low DC voltage used is 5V and the high DC voltage
is 900V . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

Figure 4.18 Example current control diagram . . . . . . . . . . . . . . . . . . . .
Figure 4.19 SVPWM voltage vectors with α − β reference . . . . . . . . . . . . .

50

51

52

55

xiv

Figure 4.20 Examining dead time in one phase leg of an inverter and how a delay
in current rise can have an eﬀect similar to additional dead time . .

Figure 4.21 With additional dead time, or delay, in top side switch and positive
. . . . . . . . . .

current ﬂow, the error in output voltage increases

Figure 4.22 SVPWM duty cycles from simulated machine with diﬀerent delay
added to high side switch of phase A . . . . . . . . . . . . . . . . .

Figure 4.23 Gate signals to high and low side phase A switches used to add delay
in turn-on to represent gate oxide degradation . . . . . . . . . . . .

Figure 4.24 VAB voltage applied with and without a 100nsec delay and the result-
ing current in phase A. With an added delay, there is an increasing
error in the current.
. . . . . . . . . . . . . . . . . . . . . . . . . . .

Figure 4.25 Normalized diﬀerence in controller-calculated duty cycles for all volt-
age vectors with delay added in high side switch of phase A . . . . .

Figure 4.26 Comparing normalized duty cycle duration with same delay added
to high side switch in all three phases where only phase A and B
currents are measured . . . . . . . . . . . . . . . . . . . . . . . . . .

Figure 4.27 Normalized duration for each SVPWM voltage vector with diﬀerent
delay added to A+ switch at diﬀerent switching frequencies . . . . .

Figure 4.28 Output voltage decrease depends on the percentage of the switching
period that the delay occupies . . . . . . . . . . . . . . . . . . . . .

Figure 5.1

After initial estimates area made, this ﬂow chart outlines the iterative
process trajectory tracking using the EKF . . . . . . . . . . . . . . .

Figure 5.2

General iterative particle ﬁlter process visually displayed where the
dots represent one particle . . . . . . . . . . . . . . . . . . . . . . .

Figure 5.3

RUL estimated using EKF compared to actual RUL from three sets
of experimental data . . . . . . . . . . . . . . . . . . . . . . . . . . .

Figure 5.4

Example cumulative distribution function used in re-sampling parti-
cles in PF process. A random value between 0 and 1 is sampled N
times from the CDF and the particle associated with that value is
added to the new population. . . . . . . . . . . . . . . . . . . . . . .

56

57

58

59

60

61

62

63

64

69

73

74

74

xv

Figure 5.5

Example trend in leakage current overshoot with peak, failure thresh-
old and estimated time of failure tEF labeled . . . . . . . . . . . . .

75

Figure 5.6

Process of estimating RUL where EKF or PF used to estimate β . .

76

Figure 5.7

Example PDF showing how conﬁdence intervals are calculated from
the distribution of tEF calculated from each algorithm . . . . . . . .

Figure 5.8

RUL estimated using EKF and PF compared to actual RUL from
two sets of experimental data with 20% conﬁdence intervals provided

Figure 5.9

RUL estimated using EKF and PF compared to actual RUL from
last 25% of actual lifetime with 20% conﬁdence intervals shown . . .

Figure 5.10 Prognosis comparison using peak-to-peak and positive peak magni-
tude from Machine 2 . . . . . . . . . . . . . . . . . . . . . . . . . .

Figure 5.11 Machine 4 RUL from data collected from oscilloscope and analog
circuit compared with the actual RUL . . . . . . . . . . . . . . . . .

Figure 5.12 RUL estimated using EKF and PF compared to actual RUL from
two sets of experimental data . . . . . . . . . . . . . . . . . . . . . .

Figure 5.13 RUL estimated using EKF and PF without accurate training data
compared to last 25% of actual RUL with 20% conﬁdence intervals
provided . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

Figure 5.14 RUL estimated using EKF and PF compared to actual RUL of Ma-
chine 1 where data sets from Machines 2 and 3 are used for training

Figure 5.15 RUL estimated using EKF and PF compared to actual RUL of Ma-
chine 5 where data sets from Machines 2 and 3 are used for training

Figure 5.16 Highlighting the diﬀerence in leakage current overshoot trends be-
tween two machines with similar ratings and diﬀerent insulation sys-
tems
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

Figure 5.17 Algorithm for the proposed prognosis algorithm where X% is 1% and
tx is the last two time steps . . . . . . . . . . . . . . . . . . . . . . .

Figure 5.18 Comparison of proposed prognosis algorithm applied to Machine 2
data when accurate training is and is not available used to tune the
algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

78

79

80

83

84

86

86

87

89

90

91

92

xvi

Figure 5.19 Comparison of proposed prognosis algorithm applied to Machine 3
data when accurate training is and is not available used to tune the
algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

Figure 5.20 Proposed prognosis algorithm applied to Machine 1 using tuning from
Machines 2 and 3, where Machine 1 degraded at a diﬀerent applied
temperature . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

93

94

Figure 5.21 Proposed prognosis algorithm applied to Machine 5 using tuning from

Machines 2 and 3, where Machine 5 had a diﬀerent insulation system 94

Figure 6.1

Eﬀect that degrading switching device has on the RUL estimates from
the EKF and PF . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

97

Figure 6.2

Eﬀect injecting a 20% decrease in overshoot magnitude has on the
RUL estimates from the EKF and PF. 20% is lower than the de-
crease that was found experimentally, and also shows a signiﬁcant
underestimation in RUL.
. . . . . . . . . . . . . . . . . . . . . . . .

Figure 6.3

Proposed technique to improve the robustness of the insulation failure
prognosis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

Figure 6.4

Vgs from degraded switching device shows increase in Miller plateau
as leakage current decreases . . . . . . . . . . . . . . . . . . . . . . .

97

98

99

Figure 6.5

Switching device applying voltage across insulation of Machine 10
shows no signiﬁcant change in Miller plateau voltage while the leakage
current overshoot rises and decays, indicating that a healthy device
and degraded insulation is present . . . . . . . . . . . . . . . . . . . 100

xvii

KEY TO ABBREVIATIONS

EKF Extended Kalman Filter

RUL Remaining Useful Life

PF Particle Filter

CI Conﬁdence Interval

WBG Wide Band Gap

HP Horse Power

HEF High Electric Field

MSE Mean Squared Error

PDF Probability Density Function

CDF Cumulative Distribution Function

FFT Fast Fourier Transform

Si Silicon

SiC Silicon-Carbide

IGBT Insulated Gate Bi-polar Transistor

MOSFET Metal-Oxide-Semiconductor Field-Eﬀect Transistor

FEA Finite Element Analysis

PD Partial Discharge

SVPWM Space Vector Pulse-Width Modulation

PWM Pulse-Width Modulation

xviii

Chapter 1

Introduction

A growing interest in electriﬁcation across industries will lead to an increase in use of electric

machines and drives. Electric machines are susceptible to a variety of diﬀerent faults [2].

The majority of faults are found to occur in the stator. Around 25% of these stator faults are

caused by degraded insulation in low-voltage machines. For high-voltage machines insulation

degradation causes around 50% of the reported failures. There are many sources of stress

that lead to short circuits and impact the lifetime of the insulation in an electric machine.

The advent of wide bandgap (WBG) devices brings a concern to the reliability of insulation

in inverter-driven machines as these devices add voltage stress to the insulation that reduces

the lifetime by an unknown amount and leads to an increase in unexpected failures.

For some applications, maintaining or improving system reliability is critical. In these

applications, sacriﬁcing reliability for electriﬁcation is not acceptable. Systems require pre-

ventative maintenance and condition monitoring tests, where the machine is removed from

operation, to avoid unexpected failures. Additional downtime for maintenance or repair in

an application that operates continuously, such as manufacturing or drilling applications, is

costly.

Online and non-invasive condition monitoring techniques are ideal as they detect degra-

dation without additional measurements while the machine is in normal operation; thus

avoiding additional downtime that an oﬄine test requires. An online technique that does

not require additional sensors or high frequency sampling is even more beneﬁcial where such

1

additional requirements can make the cost of implementation prohibitively expensive.

1.1 Objectives and contributions

As insulation degrades, short circuits form. Regardless of their classiﬁcation, all insulation

systems are susceptible to the formation of short circuits and have a ﬁnite lifetime. Shorts

can be found between two conductors, two phases or between the energized winding and

the grounded back iron. Typically, short circuits begin between two conductors and tend

to propagate quickly into a more serious, low-impedance short. Detecting an incipient fault

and taking mitigating action can avoid failure [3]; however, more advanced knowledge of

insulation degradation before a failure occurs is beneﬁcial to schedule downtime for repair or

replacement. In this work, an online technique to detect insulation degradation using mea-

surements that are already available in an inverter-driven machine control system without

high frequency sampling is proposed.

Inverters that employ WBG devices can operate at higher switching frequencies with

lower in switching losses; and for that reason, WBG devices are becoming more common

in inverter designs. However, WBG devices operate with shorter voltage rise and fall times

that increase the dV/dt applied to the terminals of the machine and lead to an increase in

the voltage stress on the insulation.

While WBG devices add stress to the insulation, these devices also degrade over their

lifetime. Switching devices, WBG devices or not, are responsible for 34% of failures in

inverter-drives [4]. Signatures in the current measurements that indicate switching device

degradation mimic the signatures produced when insulation degrades [5]. When using elec-

trical signals for detecting insulation faults or degradation, the ability to separate between

2

switching device and insulation eﬀects adds robustness to the overall reliability assessment

of the inverter-driven machine. In this work, an online technique to detect switching device

degradation in an inverter-driven machine is proposed. In addition to the proposed online

insulation and switching device degradation techniques, this work proposes a method to sep-

arate between degradation of the two components to achieve a more robust assessment of

the insulation health.

While detecting degradation can provide the operator knowledge of an unhealthy compo-

nent before it fails, a prognosis can provide an estimate of the remaining useful life (RUL).

Predicting RUL allows the operator to adequately schedule when to properly power-down

the machine before failure occurs. Accurate failure prognosis improves reliable operation of

electric machines; however, the prognosis needs to be robust to variability of applied stress

during operation and diﬀerent insulation systems. In this work, a robust, online insulation

failure prognosis technique for inverter-driven machines is proposed.

From the stated objectives, the contributions of the proposed work are:

1. A technique for online insulation failure prognosis at a reduced sampling rate and

without additional sensors

2. An online technique to detect gate oxide degradation in inverter switching devices using

voltage commands

3. A technique to separate between insulation degradation and switching device degra-

dation using current measurements

4. An insulation failure prognosis algorithm that is robust to errors in training data,

varying rates of degradation and variations in insulation systems

3

1.2 Organization

Background on insulation materials, sources of insulation degradation and a brief review

of standard insulation condition monitoring tests are provided in Chapter 2. Additionally,

an electrical model of insulation is provided for later development of a prognosis technique.

Chapter 3 outlines the proposed online, non-invasive technique to detect insulation degra-

dation. Current techniques from literature are reviewed and the advantage of the proposed

technique is highlighted. Experimental results are used to validate the proposed technique.

In Chapter 4, the online method to detect gate oxide degradation in inverter switching de-

vices is presented. Background information on gate oxide degradation, current methods

proposed for detecting gate oxide degradation and experimental results that illustrate the

proposed precursors for detecting gate oxide degradation are provided. Algorithms to esti-

mate insulation RUL are presented in Chapter 5. Each algorithm is applied to experimental

data sets of insulation under accelerated thermal degradation. Accuracy and robustness of

each algorithm is compared, and a more robust algorithm for insulation failure prognosis

is proposed. Knowledge of the eﬀects that gate oxide degradation and insulation degrada-

tion produce in the current measurements of an inverter-driven machine are combined in

Chapter 6 to propose a process to improve the robustness of the insulation failure prognosis.

Concluding remarks and notes on potential future work are provided in Chapter 7.

4

Chapter 2

Insulation Degradation

Stator insulation is categorized as conductor or groundwall insulation [6]. Conductor in-

sulation isolates the winding conductors from each other while the groundwall insulation

separates the energized conductors from the grounded case and back iron. All insulation

materials in an electric machine need to be resistant to moisture and resilient against foreign

contaminants and chemicals. However, each type of insulation has diﬀerent desired charac-

teristics. The required mechanical strength of the groundwall insulation is typically higher

than that of the conductor insulation material. Groundwall insulation is between sharp iron

edges of the stator slots and the conductors and therefore needs to resist being physically

broken down from vibration forces generated during operation [7]. Conductor insulation is

generally thinner than the groundwall insulation to ﬁt more copper in the stator slots. As

conductors in the same slot that are of the same phase do not have a large voltage diﬀer-

ence, the insulation between these conductors can be thin. This insulation is also closest to

the copper winding, which operates at the highest temperature, and therefore needs to be

thermally conductive.

Based on the requirements, some of the most common groundwall insulation materials

include mica, polyester ﬁlm and aramid paper. Common conductor insulation materials

include polyester or epoxy resins or varnishes.

Insulation materials have diﬀerent maximum temperature limits, for continuous opera-

tion, that determine the thermal class rating. Diﬀerent insulation thermal class deﬁnitions

5

Table 2.1 Thermal class designations for insulation systems

Class Designation Maximum temperature (◦C)

Material

Y
A
E

B

F

H
-
-
-

90
105
120

130

155

180
200
220
250

-
-
-

Polyester ﬁlm

Mica with shellac binder
Epoxy and polyester resin
Mica with epoxy binder
Mica with Silicon binder

-

Aramid paper

-

are provided in Table 2.1, and the common materials with maximum operating temperatures

associated with that class are listed [6]. Most of the common materials are Class B, F, or

H, where the aramid paper operates at higher temperatures.

Insulation degradation is caused by “TEAM” stresses, thermal, electrical, ambient and

mechanical [7]. Ambient stress can be caused by debris or other abrasive particles from

the environment or cooling system reaching the insulation. High humidity or oil, from the

bearings or cooling system, are also ambient stress that can chemically react with the insu-

lation material and cause degradation. Oil, for example, can penetrate mica and cause it to

separate. Moisture, when combined with an elevated temperature, causes some insulation

materials to break down. Materials within -ester groups, such as Mylar polyester, or -imide

groups, such as Kapton polyimide, are prone to hydrolysis [6, 8]. Mechanical stresses are

caused by the forces experienced in the machine during operation. These forces create vibra-

tions that can cause loose conductors to move and lead to increased wear on the insulation.

The rate at which ambient and mechanical stress degrade insulation, however, is not as well

understood.

6

Thermal stress in addition to mechanical stress accelerates degradation. Thermal cycling

causes the insulation and stator iron material to expand and contract, where the rate of

thermal expansion is diﬀerent in the two materials. This thermal expansion along with

vibrations leads to physical breakdown of insulation in the stator slots.

Thermal stress creates oxidation and cross-linking, chemical processes that occur above a

certain threshold temperature [9]. High temperature weakens chemical bonds in the insula-

tion material to the point of breaking. Oxidation occurs when oxygen joins the insulation’s

polymer chain at the broken bond. Cross-linking is when other molecules or polymers join

at the broken bond. Each process makes the insulation more brittle, and any layered insula-

tion will lose its bonding strength and deteriorate. The change in insulation lifetime due to

thermal stress is generally represented by an Arrhenius model, shown in Equation (2.1) [10].

However, at a lower temperatures, below the threshold for oxidation, no signiﬁcant thermal

degradation occurs [11].

k = Ae

Ea
RT

(2.1)

Partial discharges (PDs) are one of the most common causes of electrical stress on in-

sulation. A PD is a spark that occurs in air pockets or on the surface of the insulation.

In this process, electrons and ions bombard the insulation. For insulation made of organic

materials, the electrons and ions cause chemical bonds to break and slowly deteriorate the

insulation.

Similar to PDs, voltage surges applied on the phase terminals create electrical stress on

the insulation. Inverter pulse-width modulation (PWM) pulses create this stress, even in

low-voltage machines. The voltage pulses cause electrons from surface imperfections on the

7

copper conductors to pass into the insulation. Over time, this process causes enough bonds

in the insulation to break and the insulation to fail.

The risk for insulation failure to occur from electrical stress increases when WBG de-

vices are employed in an inverter-drive. WBG devices have a short voltage rise time which

increases the applied dV/dt to the load. Higher voltage stress caused by the WBG devices

in the inverter-drive can decrease the lifetime of the insulation by an unknown amount of

time.WBG devices can operate at higher frequencies; however, high frequency does not sig-

niﬁcantly degrade insulation [12]. While high frequency switching may be implemented with

WBG devices, the increase in terminal voltage from the quicker voltage rise degrades the

insulation more signiﬁcantly than the frequency.

Insulation material datasheets typically provide information regarding the change dielec-

tric strength at diﬀerent operating temperatures. A decrease in dielectric strength from

thermal degradation can lead to earlier failure when WBG devices used as switching devices

in an inverter-driven machine. WBG devices apply voltage with quicker rise times which

causes higher voltage stress on the insulation with lower dielectric strength. An example

plot is shown in Figure 2.1. The curve shown is from the datasheet of Nomex 410 insulation

paper, as it is a common insulation material used [7]. The decrease in dielectric strength

with temperature is logarithmic, represented by Equation (2.2).

(cid:0) 8262

273+◦C

−11.44(cid:1)

Dielectric Strength = 10

(2.2)

Voltage across the insulation, along with temperature, have a signiﬁcant eﬀect on lifetime

[12]. Changes in insulation lifetime due to voltage stress is also speciﬁed on the material

datasheet. The eﬀect of voltage stress on operating lifetime for Nomex 410 is provided in

8

Figure 2.1 Eﬀect of thermal stress on dielectric strength [1]

Figure 2.2 [1].

Figure 2.2 Eﬀect of voltage stress on time to reach 5th of 10 failures [1]

2.1

Insulation model

Insulation can be represented as an impedance, a resistance and capacitance, between two

conductors. A simpliﬁed model of groundwall insulation is shown in Figure 2.3. Along with

the dimensions of the insulation, the resistivity of the material determines the resistance and

9

the dielectric constant determines the capacitance value, shown in Equations (2.3) and (2.4).

Aε0εr

C =

d

ρl
A

R =

(2.3)

(2.4)

Figure 2.3 Simpliﬁed model of groundwall insulation between energized conductor in the
stator slot and the grounded back iron

Physical changes as insulation degrades lead to a change in its equivalent impedance.

Particulates and moisture evaporate from material when exposed to high temperatures,

resulting in an overall decrease in material, shown in Figure 2.4. This can lead to change

in capacitance, as described by Equation (2.3) where a decrease in area A decreases the

capacitance. Generally, capacitance is known to decrease as insulation degrades. However,

the capacitance will increase within 100 hours of high temperature exposure [11]. Other

sources of stress, such as partial discharges or vibrations, that cause the insulation material

to physically break also cause changes in the equivalent capacitance. The physical changes

in insulation during degradation are directly related to the capacitance.

Winding impedance is represented as a series of transmission line models as shown in

Figure 2.5. The impedance mismatch between the cables and motor winding along with

a high dV/dt voltage pulse produces a signiﬁcant reﬂected voltage at the terminals of the

10

Figure 2.4 As insulation degrades, physical changes lead to a decrease in overall area of
material around conductors

machine [13]. The reﬂected voltage is highest at the terminals and is distributed unevenly

throughout the winding, as shown in Figure 2.6.

Figure 2.5 Transmission line model of one phase winding implemented in Simulink, where
the capacitance and resistance to ground represents the insulation

Figure 2.6 Voltage measured from phase-to-ground at the terminal and near the neutral
point in a simulated inverter-driven, three-phase winding

Assuming that the applied inverter voltage can be represented by a step function, applied

11

to the transmission-line section shown in Figure 2.7, the transient response in the leakage

current can be derived.

Figure 2.7 One of the several transmission-line models, used to represent a phase winding,
used for analyzing the change in leakage current overshoot

The impedance of this model from phase to ground is derived in the Laplace domain as

Z = R1 + sL1 +

G1
sC1

G1 + 1
sC1

and can be simpliﬁed to obtain the impedance Z.

Z = R1 + sL1 +

G1

sC1G1 + 1

The current through this impedance is shown in Equation (2.7).

I =

V
Z

= V

1

R1 + sL1 +

G1

sC1G1+1

(2.5)

(2.6)

(2.7)

The denominator can be represented as a second-order polynomial, the coeﬃcients A1 and

A2 are given in Equation (2.9) an Equation (2.10) respectively.

I = V

1

s2 + A1s + A2

12

(2.8)

A1 =

R1
L1

+

1

C1G1

A2 = R2 + G1

(2.9)

(2.10)

The transfer function between the current and voltage, shown in Equation (2.8), can be

represented as a general second-order transfer function given in Equation (2.11).

ω2
0

s2 + 2ζω0s + ω2
0

=

1
Z

(2.11)

The damping term, ζ, is related to the overshoot in the transient response. Equation (2.12)

gives ζ in terms of the transmission-line impedance, derived from Equation (2.7).

(cid:33)

ζ =

+ b1

b2

(cid:32)

1

C1G1

(2.12)

(2.13)

(2.14)

where

and

b1 =

R1
L1

b2 =

√
2

1
R1G1

The damping term is inversely related to the overshoot magnitude, as shown in Figure 2.8.

Assuming that there are no other paths to ground present, the leakage current in the

system is only current through the insulation. This indicates that the insulation capacitance

and resistance are directly proportional to the overshoot magnitude in the leakage current.

The overshoot magnitude at the switching transient in the leakage current can then be used

as the signature to indicate degrading insulation.

13

Figure 2.8 Output shows varying overshoot magnitudes as ζ varies where input to system is
step response

2.2 Simulation results

A stator slot is modeled in electrostatic Finite Element analysis (FEA), Q3D extractor,

shown in Figure 2.9. Ratings for the machine and parameters of the insulation materials

used in simulation are provided in Table 2.2. The conductor coating and slot linear are

included in the model. Although the number and location of conductors in the slot are not

accurate, dimensions of the slot and conductors match the real machine. The capacitance

and conductance between the conductors and the back iron from the simulation are used

to approximate the impedance of the insulation used in the transmission-line model of one

phase of a stator winding, shown in Figure 2.5. To mimic inverter excitation, DC voltage

is applied across the transmission-line model using the voltage rise time from a MOSFET

device provided in [14].

Capacitance and resistance values of the insulation in the transmission-line model are

decreased to represent degraded insulation. Figure 2.10 shows the leakage current at the

switching transient in the case of a healthy and unhealthy insulation model. When comparing

the current waveforms, the magnitude of the overshoot in the transient response of the leakage

14

00.511.522.5300.20.40.60.811.21.41.6 = 0.2 = 0.5 = 0.8Step ResponseTime (seconds)AmplitudeTable 2.2 Parameters of the insulation machine modeled

Parameter

Voltage

Power Rating

Poles
Slots

Slot Linear Relative Permittivity (r)

Slot Linear Bulk Resistivity

Wire Coating Relative Permittivity (r)

Wire Coating Bulk Resistivity

Value

460V
1
2HP

4
24
3.5

1 × 1012
1 × 1013

3.6

Figure 2.9 One slot of random-wound induction machine with insulation included

current decreases as the insulation degrades.

Figure 2.10 Simulated voltage applied to healthy and degraded winding insulation transmis-
sion line model and leakage current responses. Insulation impedance is decreased by 50% in
this simulation.

The transmission-line model of the insulation is updated to include insulation impedance

15

100120140160180200Time ( sec)-20020406080100120140160180Voltage (V)-30-20-100102030Current (mA)Applied voltageLeakage with healthy insulationLeakage with degraded insulationbetween turns and phases [15], shown in Figure 2.11. Decreasing the capacitance and re-

sistance of the turn insulation also produces a decrease in the magnitude of the transient

overshoot, as shown in Figure 2.12. Detecting degradation of turn insulation in addition

to degradation of groundwall insulation is ideal as turn-to-turn short circuits tend to occur

before a short to ground.

Figure 2.11 Model of one phase winding with turn-to-turn insulation modeled as parallel
resistance and capacitance

Figure 2.12 Leakage current transient response shows 3.57% decrease at voltage transient in
simulation when turn-to-turn insulation impedance decreases by 50%

16

Chapter 3

Insulation Degradation Detection

Technique

There are several available standard tests that are commonly used for detecting insulation

degradation. Some of these tests along with their drawbacks are provided in Table 3.1 [16–18].

Partial discharges are quick events that require special equipment to perform high-

frequency sampling, which can be prohibitively expensive. However, measuring partial dis-

charges is commonly used for detecting insulation degradation in high-voltage machines [17],

performed oﬄine or online. PD events are diﬃcult to determine due to any noise present in

the measurements [19]. Experience is required to establish a reliable threshold to distinguish

a partial discharge event from measurement noise.

Due to the diﬃculty in accurately assessing the condition of the entire insulation system,

some maintenance protocols combine tests to better decide when to replace or remove a

machine from operation [20]. In such a protocol, tests to detect existing faults and assess

the condition of the groundwall insulation are preformed ﬁrst, followed by tests to assess the

phase-to-phase and turn-to-turn strength of the insulation. A failure threshold for these tests

is set based on experience. The frequency at which these tests are performed is also set based

on experience. While these maintenance protocols are robust methods to avoid unexpected

failures, the tests involved do not provide a failure prognosis and require a signiﬁcant amount

17

Table 3.1 List of standard insulation tests with their drawbacks

Test

Oﬄine/Online test

Disadvantage

Winding resistance

Inductive impedance

Insulation resistance

Surge test

DC high potential
AC high potential
Polarization index
Dissipation factor

Temperature monitoring

Vibration monitoring

High frequency impedance

Air gap ﬂux signature

Zero sequence voltage monitoring

Oﬄine

Oﬄine

Oﬄine

Oﬄine
Oﬄine
Oﬄine
Oﬄine
Oﬄine
Online
Online
Online
Online
Online

Only detects existing
shorts between turns
Only detects existing
shorts between turns

Dependent on

winding temperature
Can be destructive
Can be destructive
Can be destructive
Performed oﬄine
Performed oﬄine

Requires additional sensors
Requires additional sensors
Requires additional sensors
Requires additional sensors
Requires access to neutral

of data and expertise to set a failure threshold.

Several non-standard online insulation condition monitoring techniques are present in

literature. In [21], a technique that assess insulation quality using common mode current

measurements is presented. High sensitivity current transducers, as described in [2], are used

to measure phase currents. From the phase currents, the leakage current is calculated and the

dissipation factor and capacitance of the insulation are calculated. Dissipation factor is the

ratio of capacitive current to resistive current through the insulation and provides an indica-

tion of a change in the material properties. Degradation is detected from an increase in the

dissipation factor or decrease in capacitance. While this technique is demonstrated to work

well online, it requires special current transducers for accurate leakage current measurement.

A method to perform standard tests, such as Insulation Resistance and Dissipation Fac-

tor, to monitor the insulation health in an inverter-driven machine without disconnecting the

machine is available [22]. This method uses the inverter to apply voltage that generates leak-

18

age currents to assess a change in the condition of the insulation. Measuring leakage current

by sampling voltage drops over two measuring resistors at a high frequency can also provide

information regarding the insulation health [23]. Detecting degradation in groundwall and

phase-to-phase insulation separately, using Dissipation factor, resistance and capacitance,

can also be performed online in an available technique [24].

A fast Fourier transform (FFT) of the common-mode voltage and leakage current are used

to calculate the impedance of the insulation [25]. The dissipation factor and capacitance are

calculated. Results of the capacitance trends show a consistent decay from healthy values

at diﬀerent rates of degradation.

An online method to monitor insulation degradation only using current measurements is

introduced in [26]. This method samples phase current at 40M sa/sec and uses the amplitude

spectrum for detecting degradation. An FFT is calculated from healthy insulation and used

as a reference. A statistical indicator is calculated by comparing the amplitude spectrum of

each current FFT. In [27], several thermal cycles are performed and the statistical indicator

remains unchanged until failure occurs. Small changes in insulation capacitance can be

detected using this statistical indicator. This method is noninvasive and uses available

current measurements to detect insulation degradation.

These techniques are able to detect insulation degradation online in an inverter-driven

machine; however, each has disadvantages. Some techniques require additional sensors or

special, high-sensitivity sensors to detect the feature of degradation. Less-expensive sen-

sors are explored for one technique, but results in inconsistent trends in the degradation

feature [28]. High frequency sampling of current or voltage measurements is used in some

techniques, which requires more expensive equipment. Experiments used to validate the

available techniques are performed with simulated degradtion in some cases. An FFT to

19

detect insulation degradation may not the most robust feature. In addition to require high-

frequency sampling, the frequency content in the voltage and current can be aﬀected by

other sources, such as a degrading switching device [5].

3.1 Proposed technique

An online method that does not require additional sensors or high frequency sampling to

monitor insulation degradation is desired. The proposed technique meets these objectives

where the magnitude of the peak-to-peak overshoot in the leakage current is the unique

feature proposed to monitor insulation degradation. Monitoring insulation health using a

peak-to-peak magnitude is advantageous over other features, such as an FFT, as it can be

measured at a lower sampling frequency.

In an inverter-driven machine, phase currents are typically required for control. Industrial

systems that employ three-phase machines may only use two current sensor to reduce the

cost of measuring all three phase currents.

In this case, the system is assumed to have

balanced currents and the third current is calculated. However, as insulation degrades, the

machine may no longer be in balanced operation and leakage current is present.

With current in all phases measured, both the phase and leakage currents are assumed

to be available online. The sum of all phase current measurements is zero in a balanced

system. Any non-zero sum is the leakage current as shown in Equation (3.1). When all

phase currents are measured during operation, the summation of the currents provides the

leakage current measurement online.

Ia + Ib + Ic = Ileak

(3.1)

20

In proposed technique, the leakage current feature for detecting degradation is only the

peak-to-peak magnitude; analysis of the frequency content is not necessary. However, the

required sampling rate to capture the peak magnitude of the leakage current overshoot may

be signiﬁcantly high for practical online implementation. In order to ensure the peak value is

measured, Nyquist’s sampling theorem requires that the signal needs to be sampled at two-

times the frequency that is desired to be captured. Oscillation of the leakage current varies

between machines as the impedance of the insulation, which determines the leakage current

oscillation frequency, is not equal between machines. Regardless of the exact sampling

frequency required for measuring the peak in a particular machine, reducing the required

sampling rate is desired.

Reduction of the required sampling frequency using a low-cost analog circuit to measure

the overshoot magnitude is proposed. Since only the peak value is required for detecting

insulation degradation, holding this value for a longer duration allows for a sampling rate

reduction to obtain the information required to detect insulation degradation. In this work,

a circuit to follow a signal, detect when a peak is reached, and hold the peak value for a

longer duration is proposed for online detection of insulation degradation.

Thermal degradation experiments are used to validate the ability of the proposed leakage

current feature to monitor insulation health. Simulating degradation, by inserting capacitors

between turns or ground, is used to demonstrates the diﬀerence expected change in the

feature of degradation in other techniques. However, measuring the feature of degradation

using applied stress provides the overall trend expected in the feature of degradation from

healthy to failure. Measuring the trend where insulation is experiencing physical changes is

a more realistic method to evaluate the proposed feature for detecting degradation.

21

3.2 Experimental setup

Two experimental setup are developed for measuring leakage current. A simulated degrada-

tion experiment is used to demonstrate the ability to measure leakage current online using

current sensors. The second setup is to collect leakage current while insulation is thermally

degraded from healthy to failure.

3.2.1 Simulated degradation

An inverter-driven permanent magnet synchronous machine (PMSM) operating at a steady

state speed and current command is used to measure leakage current. Motor parameters

and the experimental operating point are provided in Table 3.2 and Table 3.3 respectively.

Table 3.2 Ratings of PMSM used to collect leakage current measurements

Parameter

Value

Rated power
Rated torque
Rated speed
Rated voltage
Rated current

4.8kW
45Nm
750rpm
480Vll
25A

Table 3.3 Operating point at which leakage current is measured

Operating point

Value

Speed
Current

300RPM

2.15A

In [26], insulation degradation is simulated by placing capacitors between turns in the

winding. The increase in capacitance between turns creates an increase in the insulation

capacitance. The selected PMSM has access to locations between diﬀerent turns in the phase

A winding. Table 3.4 lists the available number of turns between the diﬀerent labeled point

22

Table 3.4 Number of turns between diﬀerent labeled terminals

Terminals

Number of turns

AN1 to AN2

ANT1 to ANT2
ANM1 to ANM2
ANB1 to ANB2

120
15
14
1

in the phase A winding [29]. Leakage current is recorded without additional capacitance and

with a 15nF capacitor inserted in parallel with 14 turns, as shown in Figure 3.1.

Figure 3.1 Winding schematic with 15nF capacitor inserted across 14 turns

An oscilloscope is used to measure the three-phase currents and the leakage current is

processed from this data. The sampling rate on the oscilloscope used to measure this current

is 500M Hz.

3.2.2 Accelerated thermal degradation

Accelerated degradation tests are performed to monitor the trend in leakage current over-

shoot as insulation degrades. Experiments are performed using the stator of a three-phase

induction machine to evaluate insulation system of an electric machine. In these tests, the

23

machine is not in operation when leakage current is measured. Voltage is applied directly

over the insulation and leakage current is measure directly.

The goal of the experiment is to apply high dV/dt voltage pulses, to mimic inverter

excitation, and measure the leakage current at the switching transient while the insulation

degrades. Thermal stress is selected to accelerate the degradation of the insulation. Other

sources of stress, voltage, ambient or mechanical, are expected to decrease the equivalent

resistance and capacitance of the insulation as well; which will produce a similar decreasing

trend in leakage current overshoot. However, thermal stress is applied uniformly throughout

the winding which is desired to achieve experiments that are repeatable.

Accelerated thermal degradation is expected to cause a decrease in insulation capacitance

and resistance; however, an initial increase in overshoot magnitude is found to occur within

100 hours of thermal degradation [11]. This increase in capacitance can be caused by an

initial volume increase from thermal expansion, as the material absorbs heat, or it can also be

due to an increase in dielectric of the material due to an oxide layer forming at the interface

of the copper and insulation. Other sources of stress may not cause the initial increase in

leakage current overshoot but the decay is expected regardless of source of stress.

Tests are performed with the machine placed in an environmental chamber set to a high

temperature. The end bells, bearings, rotor and shaft are all removed and only the stator of

the machine is placed in the chamber, due to its limit size. Electrical connection can be made

to the phase leads through the side of the chamber while it is set to a high temperature, as

shown in Figure 3.2.

The experimental setup to apply voltage pulses and record leakage current is shown in

Figure 3.3. One-phase, 120Vrms AC voltage is rectiﬁed to 160V DC and a MOSFET pulses

this voltage to mimic inverter PWM excitation. The voltage is applied between one phase

24

Figure 3.2 Environmental chamber used to degrade the insulation where electrical connec-
tions can be made through the side of the chamber

lead and the case of the machine where the other two phase leads are open.

Figure 3.3 Experimental setup for measuring leakage current

The only path between phase and case is through the insulation, so the experimental

setup directly measures leakage current. An oscilloscope is used to measure leakage current

by measuring the voltage drop across a series resistor. Voltage pulses are applied continuously

and the oscilloscope is triggered to save data every ﬁve minutes for the entirety of the test.

As insulation degradation is a slow process, continuous sampling of leakage current is not

25

Table 3.5 Component operating parameters used in analog peak detector circuit

Component

Description

Op-amp slew rate

900V /µsec

Diode reverse recovery time

Capacitor

4nsec
47pF

necessary as a signiﬁcant change in insulation health does not appear instantly.

An analog peak detector circuit that consists of an op-amp with a high slew rate, a diode

and a capacitor is implemented during accelerated degradation testing, shown in Figure 3.4.

The capacitor voltage is taken as the output of the peak detector circuit. The output is

recorded and compared to the leakage current trend in order to determine if a lower sampling

rate can be used to detect insulation degradation. Details regarding this circuit are provided

in Table 3.5.

Figure 3.4 Experimental measurement setup with peak detector shown

3.2.3 Machines under test

Machines that ﬁt in the environmental chamber are selected for accelerated degradation

testing. Ratings of the selected machines are provided in Table 3.6. All of the machines are

oﬀ-the-shelf and from the same manufacturer, but are not identical. Information regarding

the insulation material in each machine is not available.

26

Table 3.6 Ratings of the machines used in accelerated thermal degradation testing

Parameter

Power
Voltage
Current

Value

1/2hp
460V
0.85A

Frequency

3450rpm

Insulation thermal class

F

Machines are from two diﬀerent batches; four are from Batch 1 and the other ﬁve are

from Batch 2. Visually, the insulation in the slot and end winding of the stator is diﬀerent

between the two batches of machines, as shown in Figure 3.5.

(a) Machine from Batch 1

(b) Machine from Batch 2

Figure 3.5 Healthy images of insulation in machines from two diﬀerent batches, where the
winding and amount of insulation between the two batches are visually quite diﬀerent

Machines 5-9, from Batch 2, have conductors with smaller diameter, more insulation

paper material separating the phases in the end winding and wedges in the slot near the air

gap. Measuring some geometric parameters show that the diameter of the conductors from

Batch 2 machines is 22.6% smaller than that of Batch 1, shown in Table 3.7. This mea-

surement is of one conductor with the insulation included, so it is unknown if the conductor

insulation is signiﬁcantly diﬀerent or if its just a smaller copper diameter that creates this

27

diﬀerence. Total slot area is also smaller in the Batch 2 machines, where the slot height is

26.7% smaller than the slot height of Batch 1 machines. With smaller conductor diameter,

there may be the same number of turns in each slot even with the reduced slot size, but that

is also unknown.

Table 3.7 Visible geometric diﬀerences between the stators and insulation in the two batches
of motors used in accelerated degradation experiments where Batch 2 motors are the higher
eﬃciency designs

Parameter

Slot height

Phase insulation thickness

Conductor diameter

Slot opening

End winding height

Batch 1 Value Batch 2 Value

16.46mm
0.25mm
0.53mm
2.74mm
35.56mm

12.07mm
0.15mm
0.41mm
2.41mm
33.25mm

These geometric diﬀerences, along with the unknown diﬀerence in the materials, cause

the insulation impedance to be diﬀerent between the machines. It can also cause the rate

at which the insulation degrades to diﬀer between machines. Even changes in insulation

properties that result from uniform applied thermal stress on identical test samples show

a wide distribution [30]. Therefor, when insulation is no identical, the resulting impedance

change between samples is likely to have a wide distribution.

3.2.4 Failure criteria

A daily visual inspection of the insulation is used to determine failure. Temperature of the

environmental chamber is lowered to room temperature every day for this inspection and is

set again to the high temperature if the insulation has not failed. Conventional insulation

tests such as the High Potential or Surge test were not performed to determine failure, as

these tests can cause signiﬁcant degradation and premature failure. The only degradation

28

desired in these experiments is from high temperatures, for the tests to be repeatable.

Since the stator is the only part placed in the environmental chamber, and the end

bells are removed, the conductors and insulation system are easily visible for inspection.

Failure is determined in each machine by visually locating conductors with missing insulation.

Figure 3.6 shows an image of the end winding of each machine where missing insulation is

found.

Figure 3.6 Images of missing insulation in Machine 1, where the lighter area in the winding
is bare copper, after failure

3.2.5 Data acquisition

Nine machine stators undergo accelerated thermal degradation of the insulation and the

trend in leakage current magnitude is obtained from each. The trend from healthy to failure

consists of periodic measurements of the overshoot in the leakage current at the voltage rise

time. The peak-to-peak value in the transient response of the leakage current, as shown in

Figure 3.7, is taken as one data point.

Measurements from the oscilloscope contain a signiﬁcant amount of noise. Wavelet de-

noising is used to remove noise to obtain a smooth waveform for analysis without adding

a phase shift, as shown in Figure 3.8. Daubechies 4 wavelet with 4 levels is selected for

de-noising.

29

Figure 3.7 Leakage current measurement on left, where zooming in and examining the leakage
current at the voltage rise on the right shows the transient oscillation and the extraction of
the leakage current overshoot, Ileak

Figure 3.8 Raw signal and signal after noise is removed from experimental leakage current
of machine with healthy insulation

After de-noising, the peak-to-peak magnitude of the overshoot in the transient response

is the feature recorded to monitor degradation. A decrease in overshoot magnitude is found

as the insulation degrades, shown in Figure 3.9.

The leakage current and peak detector output are each recorded using an oscilloscope

with a high sampling rate. Figure 3.10 shows the waveform of the applied voltage, voltage

across the in-series resistor and analog peak detector output. The peak detector output

voltage increases in magnitude after the voltage pulse. The output voltage from the peak

detector is not a DC value; rather, it increases as the voltage increases and decreases after

30

00.511.522.533.544.5Time (seconds)10-7-50-40-30-20-10010203040Current (A)Raw signalFiltered signalFigure 3.9 Insulation current overshoot at diﬀerent levels of thermal degradation where a
decrease in the peak-to-peak value is measured

the peak. As the peak detector circuit uses a capacitor on the output, the voltage level that

is output depends on the size of this capacitor. A larger capacitance hold the output voltage

longer without a signiﬁcant decay; however, the larger capacitance charges slowly which

results in a lower output voltage. Output from the peak detector shows that the increase

in magnitude after the rise in the applied voltage is held for a sustained duration, shown in

Figure 3.10, allowing for a lower sampling rate to measure the magnitude.

Figure 3.10 Output signals from the leakage current measured across an in-series resistor
and the analog peak detector

31

0.511.522.533.544.5Time ( seconds)0100200Voltage Applied (V)-10010Measured Leakage Current (V)Applied voltagePeak Detector OutputLeakage current over measuring resistorThe output of the peak detector is down-sampled at diﬀerent rates in order to ﬁnd the

sampling rate that can show the same trend as the leakage current that is used to detect

insulation degradation.

3.3 Experimental results

Simulated degradation by inserting a capacitance between produces a change in leakage

current. The peak-to-peak overshoot in the leakage current with additional capacitance

is larger than the overshoot without additional capacitance, shown in Figure 3.11. With

this measurement technique, leakage current can be measured online to detect insulation

degradation.

Figure 3.11 Experimental measurement of leakage current with and without an additional
15nF capacitor inserted between turns

The temperature and period of time that it is applied to each machine is provided in

Table 3.8. Machines 1-4 are all from the same batch and the insulation in each machine is

assumed to be similar. The same high temperature is applied to Machine 2 and 3 and the

32

012345Time ( sec)-60-40-200204060Current (A)Healthy insulation15nF insertedtime of failure is comparable. Machine 1 degraded at a lower temperature for a longer period

of time and therefore has a higher time of failure. Machines 5-9 are from a diﬀerent batch of

machines that has diﬀerences in the insulation and geometry. These diﬀerences likely cause

the signiﬁcant diﬀerence in time to failure found in Machine 4.

Table 3.8 Temperatures and duration applied for accelerated thermal degradation tests

Machine Applied temperature and duration

Machine 1

Machine 2
Machine 3
Machine 4
Machine 5
Machine 6
Machine 7

Machine 8

Machine 9

250◦C for 488 hours
260◦C for 151 hours
260◦C for 439 hours
260◦C for 513 hours
260◦C for 426 hours
260◦C for 218 hours
260◦C for 345 hours
260◦C for 329 hours
260◦C for 48 hours
250◦C for 235 hours
260◦C for 20 hours
250◦C for 174 hours

Results generally show a similar trend in leakage current overshoot magnitude, as shown

in Figure 3.12. In several machines, an initial rise in overshoot followed by a decay until

the magnitude returns to a value near the healthy overshoot magnitude is measured. The

resulting trend in overshoot magnitude from the machines is not a smooth exponential de-

cay. However, the results show that the peak-to-peak value in the leakage current transient

response can be used for detecting insulation degradation. The overall trend of reaching an

initial peak followed by an exponential decay can be used for condition monitoring.

The peak detector is implemented in one accelerated degradation experiment and Fig-

ure 3.13 shows the last 70 hours of the leakage current overshoot measured from oscilloscope

and from the output of the peak detector circuit. A comparison of sampling rate of the ana-

33

(a) Machine 1

(b) Machine 2

(c) Machine 3

(d) Machine 4

(e) Machine 5

(f) Machine 6

(g) Machine 7

(h) Machine 8

(i) Machine 9

Figure 3.12 Trend in leakage current overshoot magnitude from healthy to failure in all nine
machines

log peak detector output and the current through the in-series resistor, provided in Table 3.9,

shows a signiﬁcant reduction in the required sampling rate. A similar decay in magnitude

until failure is shown in each measurement, indicating that the peak detector circuit can be

used for detecting insulation degradation.

34

0100200300400500600Operating time (hours)203040506070Overshoot (A)0100200300400500Operating time (hours)20304050607080Overshoot (A)0100200300400500600Operating time (hours)3035404550Overshoot (A)0100200300400Operating time (hours)60708090100110120130Overshoot (A)050100150200250Operating time (hours)4050607080Overshoot (A)050100150200250300350Operating time (hours)405060708090100110Overshoot (A)050100150200250300350Operating time (hours)40455055606570Overshoot (A)050100150200250300Operating time (hours)40455055606570Overshoot (A)050100150200Operating time (hours)40455055606570Overshoot (A)Table 3.9 Sampling rate of oscilloscope to measure current across in-series resistor and to
measure output of peak detector circuit

Sampling rate Reduction

In-series resistor
Analog circuit

1.0GHz
10M Hz

-

100×

Figure 3.13 Trend in leakage current and peak detector output during testing of Machine 4

35

350360370380390400410420430Operating time (hours)60708090100110Leakage current from oscilloscope (A)0.20.40.60.811.2Output from peak detector (V)Leakage current over measuring resistorPeak detector outputChapter 4

Online Gate Oxide Degradation

Detection

There are reliability issues associated with the switching devices in an inverter-driven ma-

chine. Degradation of the gate oxide insulation layer occurs from excessive thermal or electri-

cal stress in MOSFETs and IGBT switching devices [31, 32]. As a result, the characteristics

of the turn-on voltages of these switching devices vary as degradation progresses. An online

technique to detect these turn-on transient changes in an inverter-drive is proposed. The

technique detects gate oxide degradation in switching devices without additional sensors or

high-frequency sampling.

4.1 Background

Although gate oxide degradation produces similar aﬀects in MOSFETs and IGBTs, this

work focuses on the aﬀects in MOSFETs. Gate oxide degradation occurs slowly over the

MOSFET’s lifetime and is exacerbated by high temperature or high electric ﬁeld stress [33].

In an N-type MOSFET, the device enters the on-state with a positive voltage between the

gate and source, allowing current to ﬂow, shown in Figure 4.1. Defects in the interface

between the SiO2 insulator and Si or SiC semiconductor layers, in the form of oxygen

vacancies in the silicon lattice structure, cause negatively charged ions to become “trapped”

36

in this interface, as shown in Figure 4.1b [33, 34].

(a) N-type MOSFET structure where
current ﬂows along red dotted line

(b) View of insulator and semiconduc-
tor interface where negative charges are
trapped

Figure 4.1 Illustration of physical eﬀect of gate oxide degradation in MOSFET

In Si and SiC devices, the interface between the oxide and semiconductor near the gate

is vulnerable to accumulating charged particles. However, gate oxide degradation occurs

diﬀerently in Si and SiC MOSFETs. Si MOSFETs initially accumulate charge in the oxide

layer [35]. A high positive electric ﬁeld on the gate creates oxide-trapped positive charges.

Negative charges then accumulate in the interface region. These two interactions cause the

turn-on transient characteristics, such as threshold voltage, Miller plateau voltage and time,

to decrease initially and then increase [36].

SiC MOSFETs do not accumulate charges in the oxide but in the defects in the near-

interface oxide traps [34]. This is the region near the SiC-oxide interface. Defects in this

region are well known and result in an accumulation of electrons near the gate. With only

negative charges accumulating as a positive electric ﬁeld is applied to the gate, the threshold

voltage and Miller plateau voltage are found to increase with degradation. Over time, as

the negative charge accumulates at the near-oxide interface in Si or SiC devices, the positive

voltage required for the MOSFET to enter the on-state increases thereby increasing the

threshold voltage (Vth) of the device.

Figure 4.2b shows typical MOSFET turn-on gate-to-source (Vgs) and drain-to-source

37

(Vds) voltage waveforms for a healthy device and one that has degraded gate oxide. Over

time, in both Si and SiC devices, the threshold voltage, Miller plateau voltage and duration

increase, as shown in Figure 4.2b. The change in threshold voltage causes a delay in the

current rise through the device. As the Miller plateau duration increases, the time it takes

for Vds to drop from the blocking voltage to the on-state voltage increases, which decreases

the dV/dt of the voltage applied to the load.

(a) MOSFET voltage deﬁnitions

(b) MOSFET turn-on transient voltage waveforms

Figure 4.2 Changes in turn-on transient voltages with gate oxide degradation

Inverters used for motor-drives commonly employ IGBT switching devices. IGBTs typ-

ically cannot achieve switching frequencies that MOSFETs can, but they typically have

higher power ratings than Si MOSFETs and are cheaper than SiC MOSFETs. Gate oxide

degradation also occurs in IGBTs, and similar changes in the VM P , tM P and Vth result [32].

4.2 Current Techniques

There are techniques available to detect MOSFET gate oxide degradation by measuring

characteristics of device. However, they require removing the device from the circuit or

additional sensors, invasive measurements or high frequency sampling.

38

Degradation of the gate oxide layer in Si and SiC devices can be detected through mon-

itoring the on-state resistance, threshold voltage Vth, gate or drain leakage current and

gate-to-source impedance [31, 34, 36–40]. Degradation in the gate oxide also changes some

of the characteristics in the turn-on gate-to-source voltage (Vgs) such as the level and du-

ration of the Miller plateau voltage [36]. While these characteristics can eﬀectively indicate

gate oxide degradation, they all require additional measurements and sensors that are not

typically available in an inverter-drive application.

There are techniques proposed for online detection of gate oxide degradation. In [41],

an analog circuit to monitor gate leakage current is proposed. A technique that measures

on-state resistance online using an adjustable gate driver and an additional resistor to mea-

sure drain current is also available [42]. While each method can accurately measure robust

precursors of gate oxide degradation while the MOSFET is in operation, they require ad-

ditional hardware on each individual device as well as additional sensors. The additional

requirements in these solutions may not be practical for all applications.

Current rise is discussed as a method to detect MOSFET gate oxide degradation in [43].

An analog circuit is developed to capture the gate signal and delay in current rise. This

technique, however, would require additional circuits on each inverter switch. It also uses high

frequency sampling in order to measure the delay time. In an inverter-driven machine, these

additional requirements may not be practical for all applications where cost is a signiﬁcant

factor.

39

4.3 Eﬀects of gate oxide degradation

An inverter-driven, current-controlled machine typically requires phase current sensors for

control. It cannot be assumed that voltage or temperatures sensors are available to assist

with diagnostics as these sensors are not required for operation and are not always available.

Voltages and currents on the pins of each switching device also are not typically available

as this requires an excessive number of additional sensors. It is more useful to detect gate

oxide degradation in a feature of the phase current, since these measurements do not require

additional sensors or invasive measurements. When current sensors are available on all

phases, as shown in Figure 4.3, leakage current can be estimated.

Figure 4.3 Available current measurements in an inverter-driven machine. Leakage current
to ground creates an imbalance in the three-phase currents.

Gate oxide degradation in inverter switching devices leads to a couple of changes in

features that are measurable in the phase currents. An applied voltage with high dV/dt

creates a transient response in the leakage current. The leakage current takes a path to

ground through parasitic impedance from the inverter or machine, shown in Figure 4.3.

When the dV/dt of the applied voltage decreases, the overshoot magnitude of the leakage

current transient response decreases. The circuit shown in Figure 4.4 models the current

path from one phase terminal to machine ground. As shown in Figure 4.5, the change in the

magnitude of the overshoot is found to be proportional to the change in the voltage rate of

40

rise.

Figure 4.4 Circuit used in simulation to quantify change in leakage current transient response
with voltage applied at a variable dV/dt

(a) Leakage current overshoot with high dV/dt
voltage pulse from simulation

(b) Leakage current overshoot with 3x lower dV/dt
voltage pulse from simulation

Figure 4.5 Simulation result showing decrease in applied dV/dt to a motor phase causes a
decrease in peak-to-peak overshoot in the leakage current. The green arrows indicate when
the voltage reached its ﬁnal value, showing the diﬀerence in rise time.

The other feature in the phase current that is present when MOSFET gate oxide degra-

dation occurs is a delay in the phase current rise. Switching device gate oxide degradation

leads to a increase in threshold voltage. As the gate signal rises at the same rate, the time it

takes to reach the larger threshold voltage increases, as shown in Figure 4.6. Current does

not ﬂow through the device until the threshold voltage is reached; therefore, the delay in

reaching the threshold voltage creates a delay in the rise of the current to the machine phase.

The circuit in Figure 4.7 with passive component and voltage values listed in Table 4.1

are used to demonstrate the rise delay. The load models an inductive phase winding of an

41

Figure 4.6 Gate-to-source voltage in healthy (subscript 1) and degraded (subscript 2) con-
dition. With a higher threshold voltage, the time when the MOSFET is conducting, ton,
occurs further from the initial gate signal rise

electric machine. The threshold voltage of the simulated MOSFET can be modiﬁed. The

device parasitic capacitance along with the gate resistance determine how quickly the gate

voltage rises. To more accurately model the delay time, the gate signal rate of rise is found

experimentally and used in the simulated model. The results show a 35nsec delay in the

time when the current begins to ﬂow through the load when the threshold voltage doubles

from 4V to 8V , shown in Figure 4.8.

Figure 4.7 Simulated circuit used to measure change in phase current rise as MOSFET
threshold voltage increases

4.4 Accelerated degradation test

SiC and power Si MOSFET devices are experimentally degraded to quantify changes in

the turn-on transient current features. The IRF520 power MOSFET and C3M0280090D

42

Table 4.1 Values of parameters used in simulation

Parameter

Gate resistance Rgate
Phase resistance Rload
Phase inductance Lload

Insulation capacitance CIns
Insulation resistance RIns

Value

10Ω
2Ω

20mH
250 × 10−12F
5 × 105Ω

Figure 4.8 Change in threshold voltage leading to a delay in current rise with inductive load

SiC MOSFET devices, properties of each provided in Table 4.2, are selected for accelerated

degradation testing [44, 45]. Stress is applied to the gate oxide and then the turn-on voltage

and current waveforms are captured at diﬀerent time intervals to evaluate changes. The

phase-to-case connection from a real induction machine stator is used as the load for accurate

leakage current measurement. Resistive and inductive loads are also used in order to capture

device turn-on transient characteristics such as the Miller eﬀect, threshold voltage and phase

current rise time.

4.4.1 Test bed

The devices are degraded using a High Electric Field (HEF) circuit, shown in Figure 4.9.

With this circuit, the drain and sources of the device are shorted and a voltage higher than

43

Table 4.2 Properties of the IRF520 power MOSFET and C3M0280090D SiC MOSFET

Property

VDS

RDS(on)

ID

Threshold voltage Vth range

Drain current when device in on-state

Turn-on delay
Rise time tr

Input capacitance Ciss
Output capacitance Coss

Reverse transfer capacitance Crss

IRF520 C3M0280090D

100V
0.27Ω
9.2A
2V − 4V
250µA
8.8nsec
30nsec
360pF
150pF
34pF

900V
0.28Ω
11.5A

1.8V − 3.5V

1.2mA
26nsec
10nsec
150pF
20pF
2pF

the rated gate voltage is applied between the gate and source.

Figure 4.9 HEF circuit to apply electrical stress to the gate of the MOSFET

Electrical stress is selected to degrade the MOSFETs for repeatability. The devices are

at ambient temperature during all periods of applied electrical stress. The applied voltage

across the gate and source is selected to be 65V for the power Si MOSFETs as this voltage

is found to degrade but not instantly damage gate oxide [36, 39]. A lower voltage, 37V ,

is selected to degrade SiC MOSFETs as these devices are more susceptible to gate oxide

degradation due to their typically thinner oxide layer and larger number of defects in the

oxide-semiconductor interface [34].

In addition to diﬀerent voltages used to degrade the

diﬀerent devices, the HEF stress is also applied for a diﬀerent duration when degrading each

device in order to detect changes in the turn-on transient voltages and currents.

To measure the turn-on characteristics after each interval of HEF stress, the circuit shown

44

in Figure 4.10 is used. The load in Figure 4.10a is selected as one of the three shown in

Figure 4.10b. With the switches labeled S1 closed the HEF stress is applied and with S2

closed the test pulses are applied over the desired load. The switches are used to connect the

MOSFET leads between applied stress and the load without discharging the accumulated

charge on the gate.

(a) Experimental circuit used for applying high volt-
age stress and measuring device transient character-
istics and leakage current waveforms

(b) Three load types used in measuring changes in
the turn-on transient waveforms

Figure 4.10 Experimental circuit used to apply stress and measure transient characteristics
by switching connections on MOSFET leads

The values of the voltage and passive elements used for accelerated degradation of power

Si MOSFETs are provided in Table 4.3. Experiments involving the SiC MOSFETs used

diﬀerent voltages and passive elements, provided in Table 4.4.

An oscilloscope measures the voltage drop across the series-connected resistors in the

diﬀerent loads in order to collect current measurements. The information collected with

each load type is described in Table 4.5. The probes of the oscilloscope were placed on either

side of R1, R2 and R3 to measure current when those loads are used.

Leakage current changes are assessed by using a machine as a load, shown in Figure 4.10b.

Phase A is connected to the DC voltage and the case is connected to the drain of the

45

Table 4.3 Details of components and setup information for power Si MOSFET accelerated
degradation testing

Parameter

Gate drive voltage
Gate resistance Rg

VDC on resistive and inductive loads
VDC for phase-to-case connection load

Phase resistance R1
Phase inductance L1

Resistive load R2
Series resistance R3

Oscilloscope sampling rate

Value

10V
10Ω
5V
20V
10Ω
1mH
2kΩ
0.1Ω
1Gz

Table 4.4 Details of components and setup information for SiC MOSFET accelerated degra-
dation testing

Parameter

Gate drive voltage
Gate resistance Rg

VDC on load

Resistive load R2

Value

20V
100Ω
50V
1kΩ

Oscilloscope sampling rate

500M Hz

Table 4.5 Diﬀerent load types and the features measured when the MOSFET applied voltage
across each

Load type

Measurement

Resistive-Inductive

Current rise delay

Resistive

Vth, VM P

Phase-to-case connection Leakage current, Vds

MOSFET. Any current through this path therefore is through the insulation.

4.4.2 Eﬀects of gate oxide degradation

Test pulses are applied after several one-hour intervals of HEF stress and the turn-on tran-

sient waveforms are captured. Experimental waveforms from the power Si MOSFETs are

denoised using Daubechies D4 wavelet at the 6th level, as shown in Figure 4.11, in order to

46

obtain smooth waveforms to quantify the changes in voltages and currents.

Figure 4.11 Example signals collected from the oscilloscope before and after noise was re-
moved

The Miller plateau voltage magnitude and duration change with HEF stress is shown in

Figure 4.12. A signiﬁcant increase in both precursors, the voltage and duration, result after

eight hours of gate oxide degradation.

Figure 4.12 Gate-to-source voltage waveforms for healthy device and after several periods of
HEF testing

Voltage across the device drops from the blocking voltage to the on-state voltage during

the Miller plateau. As the Miller plateau increases, the applied dV/dt decreases. The peak-

to-peak overshoot magnitude in the leakage current then decreases as the applied dV/dt

47

8509009501000105011001150120012501300Time (nsec)-2024681012Voltage (V)Raw VgsRaw VdsDe-noised VgsDe-noised Vdsdecreases, as shown in Figure 4.13.

Figure 4.13 Leakage current waveforms after several periods of HEF testing

The value of Vgs where the current sharply rises is taken to be the turn-on threshold

voltage of the device. The datasheet for the IRF520 speciﬁes a drain current of 250µA as

the current level when the threshold is reached. The voltage at the point when this current

level is reached is then taken as the threshold voltage. Figure 4.14a shows a steady increase

in Vgs level when the drain current begins to rise. Results after eight hours of HEF stress

show an increase of nearly 4 V in Vth, shown in Figure 4.14b.

(a) Threshold voltage from healthy to failure

(b) Threshold voltage in healthy condition and at
failure

Figure 4.14 Drain current vs Vgs to measure Vth

Following the increasing threshold voltage trend, the delay in phase current rise also

48

0100020003000400050006000Time (nsec)-10-8-6-4-2024Current (A)Healthy2 hours4 hours6 hours8 hours-2024681012Gate to source voltage (V)-1-0.500.511.522.53Drain current (A)10-3Healthy2 hours4 hours6 hours8 hoursTurn-on current024681012Gate to source voltage (V)-1-0.500.511.52Drain current (mA)Healthy8 hoursincreases, shown in Figure 4.15a. When plotting the drain current with the Vgs, shown in

Figure 4.15b, The gate voltage level required before current ﬂows to the load is higher after

gate oxide degradation, shown in Figure 4.15b.

(a) Current rise after several periods of HEF stress
where the solid line indicates the current level at
which the device is considered to be in the on state

(b) Current rise with Vgs to show magnitude of
gate voltage required for device to being to conduct
current

Figure 4.15 Delay in current rise after several periods of HEF testing

Quantiﬁed changes in voltage and current precursors of gate oxide degradation from the

Si power MOSFETs are provided in Table 4.6. Precursors in the voltages and currents all

show signiﬁcant change after eight hours of accelerated degradation. Threshold voltage and

delay in current rise each increased steadily after each hour of applied HEF stress while the

Miller plateau duration and leakage current overshoot did not show a signiﬁcant decrease

until several hour of degradation have passed.

Voltage and current precursors of SiC MOSFETs after accelerated degradation show a

similar trend as the Si power MOSFETs. Figure 4.16a shows the Vgs and drain current of

a healthy and degraded SiC MOSFET. Two changes are visible when examining the Miller

plateau region more closely, shown in Figure 4.16b. First, the device with degraded gate

oxide shows an increase in Miller plateau voltage while the duration of the Miller plateau

remains unchanged. Second, the current rise through the degraded device occurs later than

49

9009501000105011001150120012501300Time (nsec)00.20.40.60.811.21.41.61.82Current (A)10-3Healthy2 hours4 hours6 hours8 hoursTurn-on current050100150200250300Time (nsec)024681012Gate to source voltage (V)00.511.52Drain current (mA)Healthy VgsHealthy IdVgs after 8 hoursId after 8 hoursTable 4.6 Experimental changes in turn-on transient characteristics from Si power MOSFET

HEF applied

time (hr)

Vth

VM P

VdsFall
(nsec)

0
1
2
3
4
5
6
7
8

2.97
3.65
5.3
6.06
6.46
7.93
8.55
8.46
8.63

5.89
5.85
6.63
6.99
7.78
7.9
8.39
8.43
8.57

78.3
80
167
167
173
227
278
283
290

Ipp

9.7
11.0
10.0
10.3
11.5
6.8
5.3
5.1
4.4

I

Delay
(nsec)

10.5
28.7
66.3
79.3
103.1
114.7
136
136.7
140

the current through the healthy device. The change in current rise occurring before a change

in Miller plateau duration is found in both Si power MOSFETs and SiC MOSFETs.

(a) Current rise and Vgs from healthy and degraded
SiC MOSFET

(b) Zoomed-in view of current rise and Vgs

Figure 4.16 Current rise and Vgs from healthy and degraded SiC MOSFET where ﬁgure on
right is zoomed-in view of green box area on left plot

SiC accelerated degradation results are quantiﬁed in Table 4.7. A lower voltage is used

to degrade the SiC MOSFET and a longer duration of applied stress is required in order

to measure a signiﬁcant change in turn-on transient voltage and current precursors. As the

results show, the threshold voltage has signiﬁcantly increased to double the healthy value

and the delay in current rise shows a corresponding change.

50

Table 4.7 Table of SiC MOSFET results

HEF

applied

time
(hr)

Vth

I

Delay
(nsec)

VdsFall
(nsec)

VM P

0
2
5
8
13

4.4
4.4
4.8
5.2
8.8

10.0
10.0
10.0
12.0
20.0

14.0
14.0
12.0
14.0
16.0

6.1
6.4
8.17
9.4
9.48

Some inverter-drives may operate with a variable DC link voltage, which changes the

Vds that the device applies to the load. A Spice model, provided by Wolfspeed, is used to

evaluate the eﬀect of Vds on current rise, threshold voltage, the Miller plateau voltage. As

shown in Figure 4.17, threshold voltage change and delay are not aﬀected by magnitude of

DC voltage level. The time of current rise does not change when Vds magnitude increases,

although, Vgs is impacted.

Figure 4.17 LTSpice simulation using C3M0280090D model provided from Wolfspeed where
the low DC voltage used is 5V and the high DC voltage is 900V

51

9.91010.110.210.310.410.5Time ( sec)0246810Vgs (V)012345Current (mA)Vgs lowVgs highId lowId high4.5 Proposed technique

Delay creates an error in the current that is fed back to the controller. At steady-state

speed and torque, the error creates a deviation in the commanded voltage. Monitoring the

calculated voltage commands in the controller only requires sampling at the same rate as

the main loop of the controller; thus decreasing the required sampling rate for detecting

degraded switching devices.

A simple control diagram for a current-controlled PMSM is used to demonstrate the

eﬀect phase current error has on voltage commands, shown in Figure 4.18. The operating
speed and commanded currents i∗

q are constant.

d and i∗

Figure 4.18 Example current control diagram

Commanded voltages are calculated from PI controllers, as shown in Equations (4.1)

and (4.2), that are used to minimize the error in currents, id,e and iq,e. PI gains Kp and Ki

are tuned to achieve the desired transient and steady-state performance.

(cid:90)
(cid:90)

V ∗
d = Kpid,e + Ki

V ∗
q = Kpiq,e + Ki

52

id,edt

iq,edt

(4.1)

(4.2)

Errors in the current are calculated from the measured currents id,m and iq,m and the

commanded currents, given in Equations (4.3) and (4.4).

id,e = i∗

d − id,m

iq,e = i∗

q − iq,m

(4.3)

(4.4)

Measured currents are calculated from the three phase currents using Clarke’s and Park’s

transformations, Equation (4.5), where θ is the rotor position.

 =

iq,m

id,m

2
3

cos(θ)

cos(θ − 2π
cos(θ + 2π
3 )
3 )
sin(θ) sin(θ − 2π
3 ) sin(θ + 2π
3 )





(4.5)



ia

ib

ic

When gate oxide degradation occurs in the high-side switching device in phase A, there is a

deviation in the output current ∆ia, shown in Equation (4.6).

i(cid:48)
a = ia(1 − ∆)

(4.6)

This decrease in magnitude is added to the error in q-axis current and augments it to

produce i(cid:48)

q,e.

q −(cid:16)
d −(cid:16)

q,e = i∗
i(cid:48)

d,e = i∗
i(cid:48)

iq,m − 2
3
id,m − 2
3

(cid:17)
(cid:17)

∆iacos(θ)

∆iasin(θ)

(4.7)

(4.8)

The addition of i(cid:48)
i(cid:48)
q,e and i(cid:48)

d,e into Equations (4.1) and (4.2) creates an expression for the updated commands

a causes the updated error i(cid:48)

q,e to increase and i(cid:48)

d,e to decrease. Adding

53

q and V ∗(cid:48)
V ∗(cid:48)

d , where the updated error creates a larger Vq command and more negative Vd

command for the same steady-state operating point.

(cid:90)
(cid:90)

V ∗(cid:48)
q = Kpi(cid:48)

q,e + Ki

V ∗(cid:48)
d = Kpi(cid:48)

d,e + Ki

i(cid:48)
q,edt

i(cid:48)
d,edt

(4.9)

(4.10)

This change in Vq and Vd command is then input into the inverter modulation algorithm.

A wide variety of modulation techniques are available to apply the desired AC voltage

to the machine. Space vector PWM (SVPWM) is a common modulation strategy that uses

eight vectors, shown in Figure 4.19, to reach the desired Vα-Vβ voltage; where the inverse
Park transformation Equation (4.11) relates the d − q quantities to the α − β. Four voltage

vectors are applied for a calculated duty cycle once every switching period. In steady-state

operation, with constant voltage commands input to the SVPWM algorithm, the duty cycles

for each applied voltage vector are also constant.

 =

V ∗(cid:48)

V ∗(cid:48)

α

β

 sin(θ)

2
3

cos(θ)
−cos(θ) sin(θ)





(4.11)

V ∗(cid:48)

V ∗(cid:48)

d

q

When operating at the same steady-state point, an increase in V ∗(cid:48)
α and V ∗(cid:48)

q and decrease in V ∗(cid:48)
commands changes the SVPWM input voltage commands V ∗(cid:48)
β . Assuming a constant
DC link voltage, speed and current command, a change in the α − β voltage commands

d

produces a change in the calculated duty cycles in the SVPWM algorithm. Therefore, the

variation in duty cycles can be used to detect gate oxide degradation of the switching devices

in the inverter-drive.

Normalized calculated duty cycles are proposed for detecting gate oxide degradation in

54

Figure 4.19 SVPWM voltage vectors with α − β reference

an inverter-drive application. Unlike Vd and Vq voltage commands, the PWM duty cycle

commands are able to more easily distinguish the phase where the degraded switch is located

as the individual SVPWM vectors can indicate which phase needs higher voltage at the given

steady-state point. Duty cycles are already calculated in the controller and can be saved

without additional sensors or higher frequency sampling. The duty cycles for the steady

state operating point can also be collected and compared online.

4.6 Online detection experiment

To test eﬀects of gate oxide degradation, stress is not applied to switching devices in a

working inverter. Rather, an eﬀect that mimics a delay in phase current rise is used to

quantify the change in commanded SVPWM duty cycles.

A delay in phase current rise reduces the output current. Adding dead time to the gate

signal mimics the delay. Dead time is required in the gate signals of all inverter switches to

avoid shoot-through. High and low side switching devices of the same phase typically have

dead time added to the rising edge of the gate signals. One phase leg of a two-level inverter is

55

shown in Figure 4.20a. Figure 4.20b shows example gate signals for these switches, where td

dead time is added between the falling edge of the low-side switch and the high-side switch.

If the high-side switch is experiencing gate oxide degradation, tdelay delay time is added to

the dead time before the DC link voltage is output to phase A.

(a) One phase leg of two-level
inverter

(b) Switching signals for high and low side switches with dead
time and inserted delay

Figure 4.20 Examining dead time in one phase leg of an inverter and how a delay in current
rise can have an eﬀect similar to additional dead time

Additional delay time in when the switching device begins to conduct creates an error in

output voltage. As shown in Figure 4.21, the high side switch does not begin to conduct until

dead time passes, which leads to an error in Vout The delay caused by gate oxide degradation

leads to an application of negative voltage to Vout for an increased period of time, creating

a larger error. Normal voltage error caused by dead time can be compensated; however, a

delay due to gate oxide degradation cannot be compensated as it is unknown.

Increasing delay time in a current-controlled PMSM using SVPWM shows a measurable

change in normalized duty cycle commands, shown in Figure 4.22. The duty cycle commands

with no delay are recorded and used as baseline values. Normalized values are calculated

using Equation (4.12), where ddelay,V x is the duty cycle command from the delay in vector

Vx, x going from 0 to 7.

dN,V x =

ddelay,V x
dbase,V x

56

(4.12)

(a) Current ﬂow in undesired direction with addi-
tional dead time in top switch

(b) Voltage error that is produced by current ﬂow
in undesired direction

Figure 4.21 With additional dead time, or delay, in top side switch and positive current ﬂow,
the error in output voltage increases

As the delay in the gate signal of the high-side switch of phase A increases, the duty cycle

command for voltage vectors 1 and 2 increase. These vectors both increase voltage applied

to phase A, and this increase in duty cycle compensates for the lower voltage output caused

by MOSFET gate oxide degradation. Parameters of the simulated machine along with the

selected operating point for quantifying duty cycle commands are provided in Table 4.8.

Table 4.8 Simulated operating point for comparing calculated duty cycles with diﬀerent delay
times

Parameter or operating point

Value

Rated current

Operating current

Rated speed

Operating speed

Ld
Lq
Rs
Poles
Vdc

57

25A
20A

750RPM
300RPM
37.7mH
42.4mH

2Ω
10
680V

Figure 4.22 SVPWM duty cycles from simulated machine with diﬀerent delay added to high
side switch of phase A

Experiments are performed to quantify the eﬀects of an increase in current rise delay by

adding a speciﬁed delay to the gate signal of a switching device. Voltage commands from

the controller are collected with and without a delay to show that the duty cycles can be

used to detect gate oxide degradation.

4.6.1 Experimental setup

A current-controlled interior permanent magnet synchronous machine (IPMSM) driven by

a three-phase inverter, simlar to that shown in Figure 4.18, is used to obtain experimental

data of the eﬀects of an added delay in the gate signal of a switching device. A PMSM

current controller is implemented in Real-time LabVIEW. In this controller, only two of

the three phase currents are measured. The controller has the ﬂexibility to add a delay

as low as 25nsec in the gate signals of the switching devices. Figure 4.23 shows the gate

signals output from LabVIEW to the inverter gate driver board. The signals for A+ and
A− switches without delay show the rising and falling edges occurring at the same time.

58

01234567Voltage vector0.9511.051.1Normalized duration of each vector command10kHz fs with 100nsec delay10kHz fs with 500nsec delay10kHz fs with 1000nsec delayFigure 4.23 also shows that the controller can provide the correct delay to the gate signals
when a 500nsec delay is desired between the falling edge of A− and the rising edge of A+.

Figure 4.23 Gate signals to high and low side phase A switches used to add delay in turn-on
to represent gate oxide degradation

Change in calculated SVPWM duty cycles with an added delay is experimentally recorded

using a PMSM at the steady state operating point in Table 4.9. Ratings for the PMSM and

inverter-drive are provided in Table 4.10.

Table 4.9 Experimental steady-state operating conditions

Operating point

Value

Current command

Speed

Temperature

2.15A
300RPM
25◦C

Figure 4.24 shows a comparison of the applied voltage VAB and phase A current with

and without an added 100nsec delay to the high side switch of phase A. VAB rises with a

delay and is at the DC link voltage for a shorter duration. As a result, the phase A current

at the same point in time is lower in magnitude due to lower voltage applied to that phase.

This error in current is fed back to the PI controller, and as a result, the SVPWM duty

cycles at this steady-state operating point change.

59

Table 4.10 Parameters of PMSM used for experimental validation

Rating

Rated current
Rated speed

Phases
Slots
Poles

Value

25A

300RPM

3
12
10

Rated voltage

Switching frequency (fs)

Built-in dead time

480Vll
10kHz
300nsec

Figure 4.24 VAB voltage applied with and without a 100nsec delay and the resulting current
in phase A. With an added delay, there is an increasing error in the current.

4.6.2 Results with simulated degradation

Delays of 100nsec and 500nsec are added to the rising edge of the high side switch in

phase A. The calculated duty cycles are recorded over ten electrical cycles in steady state

operation and compared to the duty cycles at the same operating point with no added delay.

Normalized duty cycle duration of vectors V1 and V2 increase signiﬁcantly when a delay is

added to the high side switch of phase A, as shown in Figure 4.25 and in Table 4.11.

Phase C current is not measured, but instead calculated assuming no leakage current

from phases A and B measurements. To evaluate the eﬀect on the controller-calculated duty

cycles in the unmeasured phase, a delay is separately added to the high side switches of

60

050100150200Time (nsec)050100150200250300Voltage (V)1.1281.131.1321.1341.1361.1381.14Current (A)Vab no delayVab with delayIa no delayIa with delayFigure 4.25 Normalized diﬀerence in controller-calculated duty cycles for all voltage vectors
with delay added in high side switch of phase A

Table 4.11 Percent change in calculated duty cycle for each voltage vector with 100nsec and
500nsec delay added in high-side switch of phase A

100nsec delay % ∆ 500nsec delay % ∆

Voltage vector

V0
V1
V2
V3
V4
V5
V6
V7

+0.02
+0.67
+0.45
−0.55
−0.71
−0.13
−0.02
+0.02

−0.75
+9.55
+9.36
−4.77
−3.56
+1.00
+2.86
−0.75

phases B and C. A delay in phase B produces an increase in vectors V3 and V4, and a delay

in phase C results in an increase in V5 and V6, as shown in Figure 4.26 and Table 4.12.

This conﬁrms that with only two measured currents, the normalized SVPWM duty cycles

calculated in the controller can indicate the presence of gate oxide degradation in the switches

of any of the three phases.

61

01234567Voltage vectors0.940.960.9811.021.041.061.081.11.12Normalized duration of each vector command100nsec delay500nsec delayFigure 4.26 Comparing normalized duty cycle duration with same delay added to high side
switch in all three phases where only phase A and B currents are measured

Table 4.12 Percent change in calculated duty cycle for each voltage vector with 500nsec delay
added in high-side switch of speciﬁed phase

Voltage vector Phase A % ∆ Phase B % ∆ Phase C % ∆

V0
V1
V2
V3
V4
V5
V6
V7

−0.35
+7.49
+6.89
−4.28
−4.14
+0.28
−0.35

0

−0.30
−0.66
+0.96
+6.34
+6.99
−3.45
−4.65
−0.30

−0.79
−3.35
−2.21
+2.11
+3.31
+8.55
+6.62
−0.79

4.7 Detection at higher switching frequencies

To investigate the signiﬁcance of a higher inverter switching frequency, a current-controlled

PMSM driven by a two-level inverter is modeled where the delay is added to the high side

switch in phase A. Parameters of the modeled PMSM as well as the steady state operating

point are provided in Table 4.13.

At 10kHz switching frequency, a 1000nsec delay is added; and at 100kHz switching

frequency, a 100nsec delay is added. Figure 4.27 shows that in each case, the normalized

62

01234567Voltage vectors0.90.9511.051.11.151.2Normalized duration of each vector command500nsec delay in phase A500nsec delay in phase B500nsec delay in phase CTable 4.13 Simulated operating point for comparing calculated duty cycles with diﬀerent
delay times

Parameter or operating point

Value

Current
Speed

Ld
Lq
Rs
Poles
Vdc

20A

300RPM
37.7mH
42.4mH

2Ω
10
680V

change in duty cycle duration in all eight voltage vectors is similar.

Figure 4.27 Normalized duration for each SVPWM voltage vector with diﬀerent delay added
to A+ switch at diﬀerent switching frequencies

The percent change, given in Table 4.14, is similar in all vectors because the delay time

occupies the same percentage of the total switching period in each case; as shown in Fig-

ure 4.28, where the switching period is described by Equation (4.13). This causes the applied

voltage to decrease by the same percentage.

tsw =

= tp − t0

1
fsw

(4.13)

63

01234567Voltage vector0.990.9920.9940.9960.99811.0021.0041.0061.0081.01Normalized duration of each vector command10kHz fs with 100nsec delay100kHz fs with 10nsec delayWhere

tp =

1
fs

(4.14)

Table 4.14 Comparison of normalized change in duty cycles with diﬀerent delay and switching
frequency (fs) simulated

10kHz fs with 1000nsec delay % ∆ 100kHz fs with 100nsec delay % ∆

−0.25
+3.98
+3.04
−2.56
−1.28
−0.02
−0.03
−0.25

Voltage vector

V0
V1
V2
V3
V4
V5
V6
V7

−0.24
+3.93
+3.04
−2.56
−1.26
−0.01
−0.03
−0.25

Figure 4.28 Output voltage decrease depends on the percentage of the switching period that
the delay occupies

For a higher switching frequency, which is one advantage of employing MOSFET devices

in an inverter-drive, a shorter delay has a more signiﬁcant impact on the calculated SVPWM

duty cycles. Since a shorter delay is more salient when operating at higher switching frequen-

cies, earlier stages of gate oxide degradation can be detected when the switching frequency

is higher.

64

Chapter 5

Robust Insulation Failure Prognosis

Algorithm

Failure prognosis is more useful than detecting degradation or the presence of a fault. An

accurate RUL estimate provides the machine operator with knowledge of when failure will

occur so that the machine can remain in operation for a longer period of time, avoiding

unnecessary downtime, while allowing the operator to properly power-down the machine

before an unexpected failure occurs. The proposed technique to detect insulation degradation

online uses leakage current overshoot magnitude as the degradation feature.

5.1 Background

An online insulation degradation detection technique developed in [26,27] are shown to detect

insulation failure as accurately as an oﬄine test. However, this technique only indicates

when insulation is no longer healthy; it is not clear if an RUL estimate can be obtained.

In [24], the insulation capacitance, resistance and dissipation factor can be measured online

and are proposed for continuous condition monitoring. While it states that tracking the

proposed features can be used for prognosis, it is again not clear how this is achieved. A

proposed prognosis technique in [46] uses leakage current measurements and an Extended

Kalman Filter (EKF) to predict RUL. In [47], a statistical model of the insulation’s lifetime

65

uses voltage measurements in an inverter-driven machine to estimate RUL. However, the

techniques in [46, 47] both require high frequency sampling of the degradation features.

Insulation strength and the rate of degradation varies between machines. A prognosis

technique must be robust with these variations in insulation lifetime.

In [25], experimental measurements of the insulation capacitance are curve-ﬁt to an

Arrhenius model and are used to predict the decrease in capacitance. RUL is estimated

based on when the capacitance crosses a threshold that indicates failure. However, this

proposed technique requires an expensive voltage sensor. Also, it has been suggested in [11]

that the Arrhenius model for lifetime based on temperature stress is inaccurate. A prognosis

based on this model, therefore, may not be robust to unexpected variations in applied stress.

In this work, insulation failure prognosis is achieved using a Bayesian ﬁltering technique.

An EKF is one Bayesian ﬁltering technique that is capable of nonlinear state tracking.

EKF has been proposed for estimating the RUL for stator insulation [46]. Parameter and

state estimation using an EKF has been applied to power MOSFET failure prognosis [48],

bearing RUL prediction [49] and battery state of charge [50]. Prediction using the EKF only

require the value of the previous estimation, not the entire history, providing an advantage

in lower required memory space. However, the accuracy and convergence of the EKF is

highly dependent on the initial state estimates.

In [49], initial state estimates are found

from training data. Without accurate training data, where there is a signiﬁcant amount

of noise in the available measurements, the EKF may not be able to provide a suﬃciently

accurate prognosis.

Particle ﬁlters (PFs) are another Bayesian ﬁltering technique that have been demon-

strated to track noisy, highly nonlinear trajectories [51]. Parameter and state estimation

using a PF has been shown to be more robust than an EKF [52]; and their use has been

66

demonstrated literature in estimating battery state of charge [53] and RUL of batteries [51],

supercapacitors [54] and bearings [55]. While PFs can more accurately track trajectories

with high degrees of nonlinearity, they typically require a high computational cost [56].

An EKF and PF are developed and applied to the data sets from the accelerated thermal

degradation experiments. Accuracy of the RUL predictions and their robustness to error in

initial conditions, rates of degradation and diﬀerences in insulation systems are compared.

5.1.1 Extended Kalman Filter

There are several Kalman ﬁlter variants available. The EKF is an extension of the linear

Kalman ﬁlter that linearizes the state transition matrix and the output matrix in order to

use a linear system of equations for prediction.

The nonlinear system model used for the EKF is shown in Equations (5.1) and (5.2).

Here, x represent the state variables, F is the state transition matrix, w is the process noise

covariance, v is the measurement noise covariance, H is the output matrix and z is the

measured output.

xk = Fk−1xk−1 + wk−1

zk = Hkxk + vk

Matrices F and H are linearized about the current values of the states using:

Fk =

δf
δx

|xk

Hk =

δh
δx

|xk

(5.1)

(5.2)

(5.3)

(5.4)

The uncertainty matrices, M and P , are used to update the Kalman gain, K, and the

67

predicted value of the state variables, as shown in Equations (5.5) to (5.8).

Mk = Fk−1Pk−1F T

k−1 + Qk−1

Kk = MkHT

k (HkMkHT

k + Rk)−1

xk = xk + Kk(zk − Hkxk)

Pk = (1 − KkHk)Mk

(5.5)

(5.6)

(5.7)

(5.8)

Matrix R represents the measurement noise covariance and matrix Q represents the process

noise covariance.

Unknown parameters of the nonlinear state trajectory can be included as states and the

EKF can predict their true values [50]. For the nonlinear function shown in Equation (5.9),

the desired state x1 can be estimated as well as the parameters x2 and x3, if they are

unknown.

x1 = x2ex3t

(5.9)

A state matrix is then composed of all three unknowns, and the state transition can assume

that these unknown parameters are constant, as shown in Equation (5.11).

x = [x1, x2, x3]T

f (x, t) =



x2ex3t

x2

x3



(5.10)

(5.11)

The EKF process ﬁrst assumes initial estimates of the states x and the posterior estimate

68

covariance matrix, P0. The forecast error covariance matrix M is updated when a prediction

is made before an observation is received and the posterior estimate covariance matrix P is

updated after an observation is received, as shown in Figure 5.1. When observations are no

longer received, the current estimated parameters of the nonlinear function that represent

the state transition are used for predicting the future trajectory of the state.

Figure 5.1 After initial estimates area made, this ﬂow chart outlines the iterative process
trajectory tracking using the EKF

5.1.2 Particle Filter

Similar to the EKF, the PF is a Bayesian ﬁltering technique where the state transition

and measurement model can be represented in the form of state equations, as shown in

Equations (5.12) and (5.13).

xk = f (xk−1, uk) + wk

zk = h(xk) + vk

69

(5.12)

(5.13)

In the PF, state equations are represented as probability density functions (PDFs), shown

in Equations (5.14) and (5.15).

f (xk−1, uk) + wk ↔ p(xk|xk−1)

h(xk) + vk ↔ p(zk|xk)

(5.14)

(5.15)

Where the PDF in Equation (5.14) is the prior probability distribution as it predicts the

value of the state variables at the next time step before receiving a measurement. The PDF

in Equation (5.15) is the likelihood function, the probability of receiving the measurement

zk given that the states have a value xk.

From initial state estimates, a prediction of the next state is made based on the previous

observations. Then, an update is made when an observation is received. The PDF for

prediction is deﬁned using the Chapman-Kolmogorov equation, shown in Equation (5.16)

[56].

p(xk|z1:k−1) =

(cid:90)

p(xk|xk−1)p(xk−1|zk−1)dxk−1

(5.16)

This equation shows the probability of being in xk given that z1:k−1 have been observed, the

forecast PDF used for a priori prediction. This PDF is created by multiplying probability

of being in xk−1 given zk−1 by the probability of being in xk given the previous state was

xk−1, and integrating over the possible xk−1 states.

The posterior probability distribution, the PDF that represents the probability of tran-

sitioning to the present state after the current observation zk is received, is desired. To

calculate the posterior probability distribution, Bayes’ rule, shown in Equation (5.17), is

70

applied.

p(A|B) =

p(B|A)p(A)

p(B)

(cid:82) p(B|A(cid:48))p(A(cid:48))dA(cid:48)

p(B|A)p(A)

=

(5.17)

Applying Bayes’ rule to the state variables and measurements, the posterior probability

distribution is obtained in Equation (5.18).

Where

p(xk|z1:k) =
(cid:90)

p(zk|z1:k−1) =

p(yk|xk)p(xk|z1:k−1)

p(zk|z1:k−1)

p(zk|xk)p(xk|z1:k−1)dxk

(5.18)

(5.19)

The solution to Equations (5.18) and (5.19), however, is only conceptual. The PDF can-

not be evaluated analytically due to the complex, high-dimensional integral [57]. To calculate

the posterior PDF, a proposal PDF (q) is used for a priori prediction. The proposal PDF is

typically the same PDF used to represent the state transition, as shown in Equation (5.20).

q(xk|z1:k) = p(xk|xk−1)

(5.20)

From this PDF, N number of samples are taken. These samples are referred to as the

particles. When an observation is received, the distance between each particle and the

observed value is calculated. A weight is assigned to each particle, where the particles closer

to the observed value receive the higher weights. Weights are assigned using the expression

in Equation (5.21), where v is the variance of measurement noise.

wi

k =

1√
2πv

2v
|zk − xi
k|

71

(5.21)

For

i = 1 : N

(5.22)

The weights are then normalized, and the posterior PDF is created based on the weighted

particles, shown in Equation (5.23).

p(xk|z1:k) =

N(cid:88)

i=1

wi
k(xi

k − xk)

(5.23)

An updated prediction is obtained by re-sampling of the posterior PDF created from

the weighted particles. The method to weight and re-sample the particles is important, and

there are a variety of methods available in literature [56]. One typical issue is that after

re-sampling, particles with a high weight duplicate themselves several times. After a few

time steps, the particles with the highest weight may be the only one left. The diﬀerent

re-sampling methods seek to preserve diversity in the particle values.

Figure 5.2 shows a simple visual example of the PF process where the dots represent

the particles. An initial distribution of N particles, with equal weight, are obtained from q.

Then, when an observation is received, the particles are weighted based on their proximity,

and a PDF is created from this weighted distribution. Particles with the highest weight are

re-sampled to obtain N particles with equal weight again. The process then continues as a

new observation is received and the weights create an updated posterior PDF for prediction.

A high-level ﬂow chart of the PF implemented for insulation failure prognosis is provided

in Figure 5.3a. The re-sampling process is described in more detail in Figure 5.3b. Weighted

particles are input to the re-sampling algorithm. With the best particles maintained, and new

samples created, the population of particles still contains a diverse selection for searching

for more accurate parameter values while keeping the latest, most accurate particles. A

72

Figure 5.2 General iterative particle ﬁlter process visually displayed where the dots represent
one particle

cumulative distribution function (CDF) is created from the normalized, weighted particles.

An example CDF is shown in Figure 5.4.

The selected number of particles impacts the computation eﬀort in implementing a par-

ticle ﬁlter. Many particles can lead to quicker convergence and a more accurate prediction;

however, each particle is assigned a weight and re-sampled using algorithms that can be

complex in order to preserve diversity.

5.2 Prognosis

Filtering techniques are used to estimate the nonlinear model that ﬁts the trend in the

insulation degradation feature found from accelerated degradation testing. The nonlinear

73

(a) Particle ﬁlter process

(b) Particle re-sampling process

Figure 5.3 RUL estimated using EKF compared to actual RUL from three sets of experi-
mental data

Figure 5.4 Example cumulative distribution function used in re-sampling particles in PF
process. A random value between 0 and 1 is sampled N times from the CDF and the
particle associated with that value is added to the new population.

model of the degradation feature is used to estimate RUL. While the nonlinear trend in the

insulation current overshoot is obtained experimentally from thermal degradation, a similar

trend would emerge when other sources of stress degrade the insulation.

74

102030405060708090100Particle number00.20.40.60.81ProbabilityPrognosis typically requires four available pieces of information about the system or

device: a behavioral model, measurement model, degradation measurements and a failure

threshold [57]. The magnitude of the overshoot in leakage current at the switching transient

increases to a peak and then exponentially decays as the insulation ages, shown in Figure 3.12.

This trend, shown in Figure 5.5, can be used as the behavioral model.

Figure 5.5 Example trend in leakage current overshoot with peak, failure threshold and
estimated time of failure tEF labeled

Equation (5.24) can be used to describe the behavior of the overshoot in the leakage

current after the peak is reached. Here, Ileak is the measured overshoot magnitude, α is

the peak value and β is the rate of decay. Using this behavioral model, the β value can be

directly related to the changing impedance of the insulation material as it degrades.

Ileak = αeβt

(5.24)

The measurement model is used to link the degradation measurement to the behavioral

model. The healthy value is taken to be the failure threshold, shown in Figure 5.5. The pre-

dicted time of failure is the point where the behavioral model intersects the failure threshold.

The diﬀerence between the predicted time of estimated failure, tEF , and the current time is

the RUL.

75

The overall RUL estimation process is shown in the ﬂow chart in Figure 5.6. Shown

in Figure 5.6a, RUL estimation begins once the peak value in the leakage current trend is

detected.

(a) Saving healthy Ileak as failure
threshold and detecting peak value in
leakage current trend

(b) Updating β estimate in Equa-
tion (5.24) and calculating RUL from
estimated failure time tEF

Figure 5.6 Process of estimating RUL where EKF or PF used to estimate β

Figure 5.6b shows how RUL is estimated as new Ileak observations are received. In this

work, α is ﬁxed; determined from the peak value and healthy leakage current as described

in Equation (5.25).

α = ˆIleak − Ileak,healthy

(5.25)

β is estimated at every time k using either the EKF or PF algorithm.

5.2.1 Metrics of accuracy

The mean-squared-error (MSE), the coeﬃcient of determination (R2), time spent within 20%

of the true RUL (t20) and the 95% conﬁdence interval (CI) produced by the RUL estimates

are used to assess the accuracy of the predicted RUL.

76

The MSE is calculated by comparing the error between the estimated RUL (RULE) and

true RUL (RULT ) for every sample (Ns). This is used to quantify the error in the prediction,

shown in Equation (5.26).

MSE =

(cid:80)Ns

E − RULi
i=1 (RULi
Ns

T )2

(5.26)

The coeﬃcient of determination, R2, is calculated by using Equation (5.27) [58]. In this

case, the the expected model is the true RUL and the data set is the estimated RUL. A

higher R2 value indicates that the variance in the estimated RUL is low and the true RUL

is a good ﬁt for the estimated RUL.

(cid:32)

(cid:113)

R2 =

nΣ(RULE ∗ RULT ) −(cid:0)Σ(RULE)Σ(RULT )(cid:1)
E) −(cid:0)Σ(RULE)(cid:1)2(cid:113)

T ) −(cid:0)Σ(RULT )(cid:1)2

nΣ(RUL2

nΣ(RUL2

(cid:33)2

(5.27)

20% upper and lower band are placed on the true RUL. A count of the number of points of

the estimated RUL that satisfy the inequality in Equation (5.28) is calculated. The number

of samples in this count is then translated into time. Total time duration within this error

band, t20, quantiﬁes the length of time that the algorithm produced an accurate prediction.

(RULi

T × 0.8) ≤ RULi

E ≤ (RULi

T × 1.2)

For

i = 1 : Ns

(5.28)

(5.29)

A CI represents a range where the RUL is expected to lie [59].

CIs can be determined by creating a distribution of tEF values from all Ns sample points.

77

To compare the EKF and PF, a 95% CI is calculated from the estimated tEF values, shown

in Figure 5.7. This CI shows a range, an upper and lower bound, that 95% of the estimated

tEF values lie within. A more accurate RUL prediction is determined based on the CI

containing the actual time of failure and a narrower range of the CI bounds.

Figure 5.7 Example PDF showing how conﬁdence intervals are calculated from the distribu-
tion of tEF calculated from each algorithm

5.2.2 Results

Experimental data sets of the leakage current overshoot magnitude are used for comparing

RUL estimates from the EKF and PF. Data sets from Machines 2 and 3 are compared as these

machines had similar insulation systems and degraded at the same, constant temperature.

Parameters used in the EKF and PF are provided in Table 5.1. The α and β values for each

are calculated from ﬁtting the trend in the experimental data after it was collected.

Calculated RUL from each experimental data set, provided in Figure 5.8, shows that

the EKF and PF can provide a prognosis before failure occurs while the EKF shows higher

accuracy over the whole lifetime of the insulation compared to the PF. Also, the RUL

prediction is underestimated before failure occurs, which is better for avoiding unexpected

failures but leads to the premature removal from operation.

78

100200300400500600700800tEF predictions0510152025Probability DensityTable 5.1 Initial trend parameters and settings used in EKF and PF to estimate RUL for
Machines 2 and 3

Algorithm

Parameter

Machine 2 Machine 3

EKF
EKF
EKF
EKF
EKF
PF
PF
PF
PF

P (0) Uncertainty matrix

R Measurement noise covariance

Qx1 Process noise covariance
Qx2 Process noise covariance
Qx3 Process noise covariance

Number of particles

Measurement noise variance

Variance of β for initial sampling

Variance of β for re-sampling

EKF, PF
EKF, PF

α(0)
β(0)

0.01
0.1
1e−5
1e−5
1e−19
100
1

200 × 10−9
20 × 10−9
40
−7e−7

0.01
0.1
1e−5
1e−5
1e−19
100
1

200 × 10−9
20 × 10−9
15.5
−3e−7

(a) Predicted RUL from EKF and PF compared
with actual RUL from Machine 2

(b) Predicted RUL from EKF and PF compared
with actual RUL from Machine 3

Figure 5.8 RUL estimated using EKF and PF compared to actual RUL from two sets of
experimental data with 20% conﬁdence intervals provided

The RUL estimate from the PF converges closer to the true RUL after more observations

have been collected. When looking near the end of life, in Figure 5.9, the PF converges well to

the true RUL. Table 5.3 quantiﬁes accuracy towards the end of life. Close convergence to the

true RUL near the end of life is important, as that is when a decision to continue operation

or not can be critical. When degradation is just beginning, experience may suggest that the

insulation is not immediately near failure as the PF shows; rather, more measurements are

79

0100200300400Operating time (hours)0100200300400500600Remaining Life (hours)True RUL20% upper band20% lower bandPF RULEKF RUL0100200300400500Operating time (hours)0200400600Remaining Life (hours)True RUL20% upper band20% lower bandPF RULEKF RULTable 5.2 Accuracy of RUL estimated using EKF and PF for Machines 2 and 3

Machine Algorithm MSE

2
2
2
2
3
3
3
3

EKF

PF N = 50
PF N = 100
PF N = 300

EKF

PF N = 50
PF N = 100
PF N = 300

1936
35551
35612
45502
21683
67817
67820
67800

R2

93.5
0.30
0.47
0.07
90.02
40.27
40.15
40.17

t20 (hr)
204.5
5.5
25.83
24.75

0

0.02

0

0.08

CI (hr)

[371.75, 452.95]
[68.92, 433.8]

[69.7, 440]
[69.4, 435.1]
[277.7, 491.6]
[16.19, 114.2]
[16.15, 116.2]
[16.1, 115.7]

needed to improve the prediction.

(a) Predicted RUL from EKF and PF compared
with ﬁnal 25% of actual RUL from Machine 2

(b) Predicted RUL from EKF and PF compared
with ﬁnal 25% of actual RUL from Machine 3

Figure 5.9 RUL estimated using EKF and PF compared to actual RUL from last 25% of
actual lifetime with 20% conﬁdence intervals shown

Table 5.3 Accuracy of last 25% of estimated RUL using EKF and PF for Machines 2 and 3

Machine Algorithm MSE

2
2
2
2
3
3
3
3

EKF

PF N = 50
PF N = 100
PF N = 300

EKF

PF N = 50
PF N = 100
PF N = 300

97.02
300.45
361.97
337.1
811.9
2630
2633
2601

R2

98.25
98.44
97.19
96.85
93.89
11.47
10.2
12.8

t20 (hr)

CI (hr)

65
5.5
25.83
24.75

0

0.02

0

0.08

[425.2, 453]

[411.25, 433.8]

[408.9, 440]

[408.7, 442.25]
[458.19, 501, 14]
[89.57, 118.53]
[87.8, 120.1]
[88.8, 119.2]

80

350400450Operating time (hours)020406080100120Remaining Life (hours)True RUL20% upper band20% lower bandPF RULEKF RUL400450500550Operating time (hours)050100150Remaining Life (hours)True RUL20% upper band20% lower bandPF RULEKF RUL5.2.3 Computational eﬀort analysis

In the EKF, the step where the Kalman gain K is updated requires an inverse matrix

calculation. The size of the K matrix depends on the number of parameters. Therefore,

fewer states greatly reduces the computation cost. The steps involved in the EKF algorithm

and the number of calculations per time step are listed in Table 5.4.

Table 5.4 Steps required to implement the EKF

Step Number

Description

Number of operations

1
2
3
4
5

Predict next state

Update M
Update K

Update state prediction

Update P

1
1
1
1
1

As shown in Table 5.5, the number of required operations in the PF algorithm is highly

dependent on the number of particles N .

Table 5.5 Steps required in implementing the PF

Step Number

Description

Number of
operations

1
2

3

4
5
6

7

Calculate Ileak for all particles

Record observation with added noise

Find error between observation

and all particles

Assign weights to all particles

Normalize weights

Re-sample particles using weights

Create new random particles sampled
around particle with highest weight

N
1

N

N
1
N

N

Each algorithm also has diﬀerent memory requirements. For the EKF, the required arrays

and matrices that need to be saved at every time step are provided in Table 5.6. Required

memory size is also dependent on the number of state variables being estimated.

81

Table 5.6 Required memory to implement the EKF

Matrix or array

Size

Fk
3 × 3

Pk
3 × 3

Mk
3 × 3

Kk
3 × 1

Hk
3 × 1

xk
3 × 3

The PF is estimating the parameters for a curve that ﬁts well with the entire history

of the observations, and therefore the entire history is required. There are N particles that

are evaluated at every time step and each particle contains estimates for Np number of

parameters. Table 5.7 shows the required matrices for the PF algorithm.

Table 5.7 Required memory to implement the PF

Matrix

Particles

Observation history

Size
N × Np
1 × Ns

Each algorithm is implemented in a MATLAB script and executed on a PC with an

Intel Core i7-4790 CPU. The EKF shows a signiﬁcant advantage in computational eﬀort

when compared to the PF, as shown in Table 5.8. Insulation degradation, however, is a slow

process and required computation time is not a signiﬁcant concern as the degradation will

progress slower than the PF computation time.

Table 5.8 Comparison of EKF and PF computational cost with Ns = 2379

Algorithm

EKF

PF, N = 50
PF, N = 100
PF, N = 300

# of operations

per sample

Execution
time (sec)

Required values to
store in memory

5
252
502
1502

1.65
7.66
14.06
32.14

42
2529
2679
3279

When estimating RUL using the PF, the number of particles is varied to ﬁnd if there

is an optimal number of particles for this application. No signiﬁcant impact on accuracy

82

is found when the number of particles is above 100. As increasing the number of particles

can have a signiﬁcant impact on computational cost, shown in Table 5.8, the insigniﬁcant

increase in accuracy with more than 100 particles is not a favorable trade-oﬀ.

5.2.4 Reduced sampling rate

To reduce the high sampling rate required for an RUL estimate, the peak detector output

needs to provide a similar trend for a similar prognosis. The peak detector used, however,

only captures the magnitude of the positive peak in the transient response, not the peak-

to-peak magnitude. In order to ﬁnd if the positive peak magnitude is suﬃcient, an RUL

estimate using an EKF with the trend in the positive peak in the leakage current measured

from the oscilloscope is compared to the RUL estimate using the peak-to-peak magnitude.

Figure 5.10 shows that a prognosis with similar accuracy can be calculated using just the

magnitude of the positive peak. The output of the peak detector, therefore, should be able

to provide an accurate prognosis.

Figure 5.10 Prognosis comparison using peak-to-peak and positive peak magnitude from
Machine 2

Trends in the peak-to-peak magnitude, positive peak magnitude and analog circuit output

83

050100150200250300350400450Operating time (hours)0100200300400500600Remaining Life (hours)RUL using peak-to-peak valuesRUL using positive peakActual RULhave a comparable rate of decrease towards the end of life, as shown in Table 5.9. The rate of

decrease from the peak detector is similar to that of the positive peak value. RUL estimates

using the trends from the peak-to-peak overshoot and the peak detector output result to be

similar, as shown in Figure 5.11. Therefore, the output of the peak detector circuit, sampled

at a 10x lower frequency, is suﬃcient for insulation failure prognosis.

Table 5.9 Comparison of the rate of decrease from the peak-to-peak trend, positive peak
trend and analog circuit output trend for the last 70 hours of life from Machine 4

Peak-to-peak magnitude Positive peak magnitude Peak detector output

10.9 × 10−3

7.9 × 10−3

7.1 × 10−3

Figure 5.11 Machine 4 RUL from data collected from oscilloscope and analog circuit com-
pared with the actual RUL

5.3 Evaluation of robustness

There are many sources of stress on insulation and any variation between two similar ma-

chines can cause the insulation lifetime to vary signiﬁcantly. The EKF is known to diverge

from the true value when measurement data has a signiﬁcant amount of noise [49]. Three

84

360370380390400410420Operating time (hours)01020304050607080Remaining Life (hours)RUL from analog circuitRUL from leakage current over measuring resistorActual RULcomparisons are performed to examine robustness of the RUL prediction from the EKF and

PF:

1. Errors due to inaccurate training data

2. Varying rate of degradation

3. Variation in insulation system

The diﬀerence in accuracy of the resulting RUL predictions is compared to evaluate which

algorithm is more robust to these three scenarios.

5.3.1 Inaccurate training data

With training data available, it cannot be assumed that all insulation will follow the same

degradation trend. After adding an error in the initial parameter estimates, again the EKF

provides a more accurate RUL estimate over the entire lifetime as shown in Figure 5.12. The

accuracy of the EKF is lower than the accuracy of the RUL produced when accurate training

data is available, shown in Table 5.10. The accuracy of the RUL from the PF, however, is

similar to the accuracy of the RUL estimate when accurate training data is available.

Table 5.10 Accuracy of RUL estimated using EKF and PF without using accurate training
data for Machines 2 and 3

Machine Algorithm MSE

2
2
2
2
3
3
3
3

EKF

PF N = 50
PF N = 100
PF N = 300

EKF

PF N = 50
PF N = 100
PF N = 300

3083
35774
36331
35259
2290
67817
67806
67797

R2

98.35
0.35
0.08
0.414
3726
40.9
41.64
39.9

t20 (hr)
196.8
4.7
20.7
30.83
194.25

0

0.02

0

CI (hr)

[460.5, 539.29]
[68.58, 431.2]
[69.75, 434.58]
[68.8, 435.92]
[391.96, 563.37]

[16.2, 115]

[16.11, 115.45]
[16.31, 115.2]

85

(a) Predicted RUL from EKF and PF compared
with actual RUL from Machine 2 when inaccu-
rate data used for training

(b) Predicted RUL from EKF and PF compared
with actual RUL from Machine 3 when inaccu-
rate data used for training

Figure 5.12 RUL estimated using EKF and PF compared to actual RUL from two sets of
experimental data

Towards the actual end of life, the PF is more accurate, shown in Figure 5.13. The EKF

now overestimates RUL towards the end of life for Machines 2 and 3. Overestimation is

undesired in applications where safety is critical as it leads to unexpected failures.

(a) Predicted RUL from EKF and PF compared
with last 25% of actual RUL from Machine 2
when inaccurate data used for training

(b) Predicted RUL from EKF and PF compared
with last 25% of actual RUL from Machine 3
when inaccurate data used for training

Figure 5.13 RUL estimated using EKF and PF without accurate training data compared to
last 25% of actual RUL with 20% conﬁdence intervals provided

Comparing the accuracy when inaccurate training data is used, in Table 5.11, to the

accuracy in Table 5.3 shows that the EKF has a signiﬁcant decrease in accuracy towards the

86

0100200300400Operating time (hours)0100200300400500600Remaining Life (hours)True RUL20% upper band20% lower bandPF RULEKF RUL0100200300400500Operating time (hours)0200400600Remaining Life (hours)True RUL20% upper band20% lower bandPF RULEKF RUL350400450Operating time (hours)020406080100120Remaining Life (hours)True RUL20% upper band20% lower bandPF RULEKF RUL400450500550Operating time (hours)050100150Remaining Life (hours)True RUL20% upper band20% lower bandPF RULEKF RULTable 5.11 Accuracy of last 25% of RUL estimated using EKF and PF without accurate
training data for Machines 2 and 3

Machine Algorithm MSE

2
2
2
2
3
3
3
3

EKF

PF N = 50
PF N = 100
PF N = 300

EKF

PF N = 50
PF N = 100
PF N = 300

1895
341.5
262

320.25
2217
2628
2626
2615

R2

99.63
98.44
97.99
98.04
92.8
9.0
11.1
12.2

t20 (hr)

CI (hr)

0
4.7
20.7
30.83

0
0

0.02

0

[478.1, 486.9]
[408.25, 433.14]
[413.6, 435.75]
[410.25, 437.7]

[538.2, 569]

[88.08, 118.78]
[89.3, 118.45]
[88.7, 119.3]

end of life while the accuracy of the PF is maintained.

5.3.2 Varying rate of degradation

In this comparison, the rate of degradation that produced the training data is diﬀerent than

the rate used to degrade the insulation of this machine. As shown in Figure 5.14, when

insulation in machines is degraded at diﬀerent temperatures, the PF is more robust and

shows a more accurate RUL estimate over the entire lifetime and towards the end of life.

(a) Predicted RUL from EKF and PF compared
with actual RUL from Machine 1

(b) Predicted RUL from EKF and PF compared
with actual RUL from Machine 1, last 50% of life

Figure 5.14 RUL estimated using EKF and PF compared to actual RUL of Machine 1 where
data sets from Machines 2 and 3 are used for training

87

0200400600Operating time (hours)0200400600800Remaining Life (hours)True RUL20% upper band20% lower bandPF RULEKF RUL400500600700Operating time (hours)0100200300Remaining Life (hours)True RUL20% upper band20% lower bandPF RULEKF RULOver the entire lifetime, the RUL estimate from the EKF has a lower MSE than the PF,

quantiﬁed in Table 5.12. However, towards the last half of the insulation lifetime, the PF

is signiﬁcantly more accurate, shown in Figure 5.14b. In this case, where training data is

obtained from machines degrading at a higher temperature, the PF is more robust to provide

an accurate prognosis towards the end of life.

Table 5.12 Accuracy of RUL estimated using EKF and PF for Machines 1 over the whole
lifetime and the last 50% of life, where data sets from Machines 2 and 3 are used for training

% of lifetime for calculation Algorithm MSE

100%
100%
100%
100%
50%
50%
50%
50%

EKF

PF N = 50
PF N = 100
PF N = 300

EKF

PF N = 50
PF N = 100
PF N = 300

29734
33804
33443
33121
21544
4019
3930
3924

R2

86.15
51.54
50.72
50.87
69.3
91.42
91.68
91.48

t20 (hr)
24.75
55.91
55.17
65.5

0

55.91
44.7
55.25

CI (hr)

[413, 608.13]
[190.41, 608.6]
[184.7, 611.93]
[184.4, 616.6]
[435.3, 624.38]
[544.2, 624.38]
[542.8, 625.7]
[538.9, 624.38]

5.3.3 Diﬀerent insulation system

When the algorithm is trained using data from a machine with a diﬀerent insulation system

and then applied to predict RUL, the PF is again more robust. The PF shows a underes-

timated RUL initially and converges close to the true RUL while the RUL estimate from

the EKF does not converge within 20%, shown in Figure 5.15. Towards the end of life,

shown in Figure 5.15b, the EKF is signiﬁcantly overestimating RUL while the PF slightly

underestimates.

In this case, the PF is more accurate than the EKF not only towards the end of life, but

over the whole lifetime in Table 5.13.

88

(a) EKF predicted RUL compared with actual
RUL from Machine 5

(b) Predicted RUL from EKF and PF compared
with actual RUL from Machine 5, last 60% of life

Figure 5.15 RUL estimated using EKF and PF compared to actual RUL of Machine 5 where
data sets from Machines 2 and 3 are used for training

Table 5.13 Accuracy of RUL estimated using EKF and PF for Machine 5 over the whole
lifetime and the last 60% of life, where data sets from Machines 2 and 3 are used for training

% of lifetime for calculation Algorithm MSE

100%
100%
100%
100%
60%
60%
60%
60%

EKF

PF N = 50
PF N = 100
PF N = 300

EKF

PF N = 50
PF N = 100
PF N = 300

15176
4862
4830
4790
8146
382.3
328.6
347.4

R2

97.8
51.91
52.29
51.91
98.2
85.5
90.5
88.2

t20 (hr)

CI (hr)

0

3.25
0.08

0
0
0
0
0

[299.3, 401.42]
[59.75, 208.1]
[58.1, 208.41]
[57.58, 201.92]
[298.8, 328.3]
[191, 215.5]
[188.8, 215.5]
[192.2, 215.5]

5.4 Failure threshold

The trend in leakage current overshoot from the machine with a diﬀerent insulation system,

Machine 5, is diﬀerent than the trend from the other three data sets used to compare

robustness of the algorithms. Data sets from the other three machines follows a trend

similar to what is shown in Figure 5.16a: an initial rise from a healthy value to a peak,

and then an exponential decay back down near that initial healthy value. Therefore, the

failure threshold is set to be when the decaying trend crosses the value of the healthy, initial

89

050100150200Operating time (hours)050100150200250300Remaining Life (hours)True RUL20% upper band20% lower bandPF RULEKF RUL100150200Operating time (hours)050100150200Remaining Life (hours)True RUL20% upper band20% lower bandPF RULEKF RULovershoot values.

The initial overshoot values from Machine 5, however, are larger in magnitude than the

overshoot values towards the end of life, shown in Figure 5.16b. This results in a signiﬁcant

underestimation of lifetime when using the same failure threshold that is used for the other

three machines.

(a) Labeled trend in leakage current overshoot
from Machine 2

(b) Labeled trend in leakage current overshoot
from Machine 5

Figure 5.16 Highlighting the diﬀerence in leakage current overshoot trends between two
machines with similar ratings and diﬀerent insulation systems

Instead of using the initial value of overshoot as the failure threshold for Machine 5,

the failure threshold is set by assuming a percent decrease from the peak overshoot value.

Overshoot trends from the other three machines show that the ﬁnal overshoot value is 50%

lower in magnitude than the peak value, on average. Therefore, for Machine 5, the failure

threshold is set to half of the value of the peak in the overshoot magnitude trend.

In this work, leakage current overshoot trends show an initial increase followed by an

exponential decay until insulation fails. This initial increase may be due to the applied

thermal stress used to accelerate the degradation, as it is also observed in [11]. Under normal

operating conditions, this quick increase may not be present, if insulation is degrading from

90

other sources of stress. Because of this, setting the failure threshold as a percentage of the

peak value in leakage current overshoot may be more accurate as the exponential decay is

expected to occur as the insuation capacitance decreases with any type of applied stress.

5.5 Proposed prognosis algorithm

As shown in Figures 5.13 to 5.15, the PF gives early false failure indication and the EKF

is not always accurate towards the true end of life. In order to provide a prognosis that is

more accurate and robust, an algorithm that combines the EKF and PF is proposed.

The algorithm, shown in Figure 5.17, will use the output of the EKF until the PF

converges. Convergence is when the current PF RUL estimate is within 1% of the previous

two estimates. When the PF converges, there may be a large diﬀerence in the EKF and PF

estimates. To avoid a sudden step in RUL estimate when the algorithm switches between

the EKF and PF, the estimate is slowly ramped between the EKF and PF.

Figure 5.17 Algorithm for the proposed prognosis algorithm where X% is 1% and tx is the
last two time steps

As shown in Figure 5.18, the proposed algorithm provides an accurate RUL estimate

over the entire lifetime even with inaccurate training data. The RUL estimate is accurate,

91

within 20% of true RUL, at the initial prediction and remains accurate, 40% of true RUL,

within the last 100 hours of the end of life without a signiﬁcant overestimation, as shown in

Table 5.14.

Figure 5.18 Comparison of proposed prognosis algorithm applied to Machine 2 data when
accurate training is and is not available used to tune the algorithm

Table 5.14 Accuracy of RUL estimated using EKF, PF and proposed algorithm for Machines
2 with and without accurate training data

Algorithm Accurate Training Data? MSE

EKF
EKF

PF N=100
PF N=100
Proposed
Proposed

Yes
No
Yes
No
Yes
No

1936
3083
35612
36331
2103
3413

R2

93.5
98.35
0.3
0.08
95

94.63

t20 (hr)
204.5
196.83

5.5
20.7
162.08
123.0

CI (hr)

[371.75, 452.96]
[460.5, 539.29]
[68.9, 433.8]
[69.75, 434.6]
[20.8, 398.4]
[18.1, 491.3]

As shown in Figure 5.19, the proposed algorithm does not always show a signiﬁcant

improvement when inaccurate training data is used for tuning. Although the PF does not

converge until near the time of failure, the overestimation towards the time of failure is

reduced. Shown in Table 5.15, the resulting RUL estimate is not signiﬁcantly improved.

In the case where training data is obtained from machines that degraded at a diﬀerent

rate, the proposed RUL estimation algorithm results in signiﬁcant improvement in accuracy

92

0100200300400Operating time (hours)0100200300400500600Remaining Life (hours)True RUL20% upper band20% lower bandRUL with trained proposed algorithmRUL with untrained proposed algorithmFigure 5.19 Comparison of proposed prognosis algorithm applied to Machine 3 data when
accurate training is and is not available used to tune the algorithm

Table 5.15 Accuracy of RUL estimated using EKF, PF and proposed algorithm for Machines
3 with and without accurate training data

Algorithm Correct Training? MSE

EKF
EKF

PF N=100
PF N=100
Proposed
Proposed

Yes
No
Yes
No
Yes
No

21683
3726
67820
67806
21684
3522

R2

90.0
91.75
40.15
41.64
90.07
90.86

t20 (hr)

CI (hr)

0

[277.7, 491.6]

194.25

[392, 563]

0

0.02

0
197

[16.15, 116.2]
[16.11, 115.45]

[10.81, 236]
[12.87, 367.7]

over the entire lifetime compared to the estimates in Figure 5.14a, shown in Figure 5.20.

Applying the proposed algorithm over the whole lifetime results in a lower MSE and higher

t20, shown in Table 5.16.

Table 5.16 Accuracy of RUL estimated using EKF, PF and proposed algorithm for Machines
1 using data sets from Machines 2 and 3 as training

Algorithm MSE

EKF

PF N=100
Proposed

29734
33443
16628

R2

86.15
50.72
85.7

t20 (hr)
24.75
55.17

80

CI (hr)

[413, 608.1]
[184.7, 611.9]
[1.89, 445.49]

Applying the proposed prognosis algorithm using training data collected from machines

93

0100200300400500Operating time (hours)0200400600Remaining Life (hours)True RUL20% upper band20% lower bandRUL with trained proposed algorithmRUL with untrained proposed algorithmFigure 5.20 Proposed prognosis algorithm applied to Machine 1 using tuning from Machines
2 and 3, where Machine 1 degraded at a diﬀerent applied temperature

with a diﬀerent insulation system, shown in Figure 5.21, also shows an improvement in

the accuracy of the RUL estimate compared to the those obtained in Figure 5.15a. The

MSE obtained is comparable to the results of the PF, but the t20 and R2 are signiﬁcantly

improved, shown in Table 5.17.

Figure 5.21 Proposed prognosis algorithm applied to Machine 5 using tuning from Machines
2 and 3, where Machine 5 had a diﬀerent insulation system

The proposed algorithm is shown to improve accuracy of the RUL estimate over the entire

life of the insulation. Compared to the EKF and PF algorithms, the proposed algorithm

generally results in a lower or comparable MSE while also resulting in a high R2 and t20.

94

0200400600Operating time (hours)0200400600800Remaining Life (hours)True RUL20% upper band20% lower bandRUL with proposed algorithm050100150200Operating time (hours)050100150200250300Remaining Life (hours)True RUL20% upper band20% lower bandRUL with proposed algorithmTable 5.17 Accuracy of RUL estimated using EKF, PF and proposed algorithm for Machines
5 using data sets from Machines 2 and 3 as training

Algorithm MSE

EKF

PF N=100
Proposed

15176
4830
5800

R2

97.8
52.29
80.1

t20 (hr)

CI (hr)

0

0.08
15.25

[299.3, 401.42]
[58.07, 208.4]
[0.7, 387.7]

The proposed algorithm also maintains robustness as it uses the output of the PF towards

the true end of life. It is shown to converge close to the true RUL when training data is

inaccurate, obtained from insulation degraded at a diﬀerent rate or obtained from a machine

with a diﬀerent insulation system.

95

Chapter 6

Fault Separation

Gate oxide degradation, leads to a signiﬁcant decrease in leakage current overshoot. This

eﬀect may cause underestimation of insulation RUL. Degradation of switching devices has

also been shown to cause a change in the harmonic content of the current at the switching

transient [5]. The insulation degradation detection method developed in [26, 27] relies on

harmonic content at the switching transient to monitor insulation health. As a result, these

two features, the overshoot magnitude and harmonic content in the current, can only be

used to monitor insulation health with an assumption of healthy inverter switching devices.

A decrease in the overshoot magnitude is observed when gate oxide degradation occurs

in a MOSFET, Tables 4.6 and 4.7. Such a decrease in overshoot magnitude results in a

signiﬁcant underestimation of RUL, shown in Figure 6.1.

The change in overshoot magnitude from gate oxide degradation may vary across dif-

ferent switching devices; however, any change in overshoot magnitude that is caused by

the switching device will lead to a signiﬁcant underestimation insulation of RUL, shown in

Figure 6.2.

6.1 Proposed technique for fault separation

An algorithm to improve robustness of the insulation RUL estimate checks if the change in

overshoot is due to insulation degradation or switching device gate oxide degradation prior

96

(a) Comparing RUL estimate using EKF when
a 30% decrease in overshoot magnitude injected
into the data set from Machine 2 to mimic the
eﬀect of a degraded switching device

(b) Comparing RUL estimate using PF when a
30% decrease in overshoot magnitude injected
into the data set from Machine 2 to mimic the
eﬀect of a degraded switching device

Figure 6.1 Eﬀect that degrading switching device has on the RUL estimates from the EKF
and PF

(a) Comparing RUL estimate using EKF when
a 20% decrease in overshoot magnitude injected
into the data set from Machine 2 to mimic the
eﬀect of a degraded switching device

(b) Comparing RUL estimate using PF when a
20% decrease in overshoot magnitude injected
into the data set from Machine 2 to mimic the
eﬀect of a degraded switching device

Figure 6.2 Eﬀect injecting a 20% decrease in overshoot magnitude has on the RUL estimates
from the EKF and PF. 20% is lower than the decrease that was found experimentally, and
also shows a signiﬁcant underestimation in RUL.

to RUL estimation, as shown in Figure 6.3. It also calculates an RUL estimate that is robust

to error in initial estimates, change in rate of degradation or change in the insulation system.

When a decrease in leakage current is observed, an additional check of the SVPWM duty

cycles is made in order to determine if the inverter switching devices are contributing to the

leakage current, shown in Figure 6.3a. If not, it is assumed that the insulation is degrading

97

and the proposed prognosis algorithm is executed to predict RUL. If there is a signiﬁcant

change in SVPWM duty cycles, it is assumed that a switching device is degrading and the

insulation is assumed to remain in its current health state.

(a) Flow chart for separating between switching
device and insulation degradation

(b) Logic to determine when to switch to PF for
RUL estimation, represented by the box in red
in Figure 6.3a

Figure 6.3 Proposed technique to improve the robustness of the insulation failure prognosis

RUL estimation in the proposed algorithm is performed only using the overshoot mag-

nitude of the leakage current, SVPWM duty cycles are only used to conﬁdently determine

if the change in leakage current overshoot is due to switching device degradation or not.

Results from accelerated degradation tests show that overshoot magnitude changes with

both gate oxide degradation and insulation degradation but only gate oxide degradation

produces a change in SVPWM duty cycles. Results show a decrease in the Miller plateau

voltage and in the leakage current overshoot as a MOSFET experienced accelerated gate

oxide degradation, shown in Figure 6.4. If there is a decrease in leakage current due to the

gate oxide degradation, an increase in the Miller plateau voltage occurs. The increase in

threshold voltage occurs as the Miller plateau voltage increases, which leads to the increase

in current rise delay and change in SVPWM duty cycles.

Accelerated insulation degradation experiments are performed using a switching device

98

(a) Leakage current overshoot where insulation
impedance remained constant and switching de-
vice degraded

(b) Vgs from degraded switching device shows
increase in Miller plateau

Figure 6.4 Vgs from degraded switching device shows increase in Miller plateau as leakage
current decreases

to pulse DC voltage across the insulation. The resulting leakage current overshoot shows

a rise followed by an exponential decay. During an experiment, the Vgs of the switching

device is recorded to show that the leakage current decreases and no signiﬁcant increase

in the Miller plateau voltage occurs, shown in Figure 6.5. This indicates that insulation

degradation does not aﬀect the switching device characteristics and does aﬀect the leakage

current. Therefore, using the proposed change in SVPWM duty cycles to detect switching

device degradation is a robust technique to determine if the switching device contributes to

any change in the leakage current overshoot.

99

(a) Leakage current overshoot trend from Ma-
chine 5 as thermal stress is applied to the insula-
tion

(b) Measured Miller plateau voltage in Vgs of the
device that is applying voltage to the insulation
of Machine 5 as it experiences accelerated ther-
mal degradation

Figure 6.5 Switching device applying voltage across insulation of Machine 10 shows no signif-
icant change in Miller plateau voltage while the leakage current overshoot rises and decays,
indicating that a healthy device and degraded insulation is present

100

Chapter 7

Conclusion and Future Work

Stator insulation prognosis can be eﬀectively performed using leakage current overshoot

magnitude as a feature of insulation degradation and a combination EKF and PF to predict

RUL. To improve conﬁdence in the RUL estimate, insulation degradation can be isolated

from inverter switching device degradation even though degradation of each component

results in a similar change in leakage current overshoot.

Techniques proposed in this work for detecting insulation degradation and switching

device degradation can each be performed online, at a reduced sampling rate and with-

out additional sensors. With current sensors on all phases, leakage current is shown to

be measurable online. An analog peak detector circuit is implemented during accelerated

degradation testing and the output shown to provide suﬃcient information for an insulation

failure prognosis at a reduced sampling rate. Gate oxide degradation in inverter switching

devices is detectable using features in the leakage or phase currents. In a current-controlled

machine, the increase in phase current delay caused by gate oxide degradation is fed back

to the controller and results in a change in voltage commands. As voltage commands are

available in the controller, online detection of gate oxide degradation is possible without

additional sensors while only sampling at the controller’s main calculation loop rate.

As switching device degradation and insulation degradation lead to a similar decrease in

the leakage current overshoot, monitoring the steady-state voltage commands is proposed

to distinguish when switching device degradation is present. Accelerated degradation of

101

insulation and switching devices shows that only switching device degradation produces

a change in the commanded voltages while insulation degradation does not. Separating

between these two faults decreases the likelihood of a false positive detection of insulation

failure.

Leakage current overshoot trends collected from accelerated insulation degradation are

used to develop a prognosis technique. Applying the PF to these trends resulted in an early

indication of insulation failure while the RUL prediction from the EKF is not always accurate

towards the end of life. Therefore, an algorithm that uses the prediction of the EKF initially

and then transitions to use the PF is proposed for RUL estimation. The proposed algorithm

is shown to be robust to varying rates of degradation, variations in insulation system and

inaccurate training data. Using the proposed degradation detection and fault separation

techniques, an online insulation failure prognosis is possible and can be made more robust

with the proposed RUL prediction algorithm.

Future work includes developing an online switching device failure prognosis method. A

change in SVPWM duty cycle commands is proposed as the feature of degradation. Further

work can use this feature to develop a protocol for RUL estimation which includes deﬁning

the behavior model, measurement model and failure threshold that are required for progno-

sis. The turn-oﬀ transient characteristics of the switching device can also be explored for

condition monitoring and prognosis.

Accelerated degradation tests of electric machine insulation can be improved. This work

used a high temperature to degrade insulation. Other works, however, discuss how insulation

properties degrade diﬀerently when at diﬀerent temperatures and that the Arrhenius model

is not true for all materials [11]. Changes in the chemical composition and how it aﬀects

the insulation properties are unknown, but these changes lead to a change in the electrical

102

impedance of the insulation material. Detailed rationale for the connection between applied

stress and the change in the electrical model for insulation degradation is also future work.

Additional future work includes improving the thermal degradation experimental setup

and procedure. Testing relied on visual inspection of insulation to detect defects; however,

visual inspection could not determine if insulation is missing has defects within the coil

bundles or inside the stator slots. Only the end winding conductors are visible in the current

test setup. Insulation could have failed well before failure was determined, and the defect

was not noticed. As a result, the trend in insulation degradtion may not accurately show

the true trend in leakage current overshoot. Accurate information regarding the design of

the machines used for testing including insulation materials used can also be used to help

justify the diﬀerences in insulation lifetime.

Future work also includes improving the prognosis algorithm.

In this work, the RUL

estimate is taken from the EKF until the PF has converged. Other techniques to determine

which algorithm trust can be developed by assigning weights or a Maximum Likelihood

Voting algorithm [60]. The EKF and PF themselves can also be improved upon as there are

a number of variations of these algorithms that can be implemented to further improve the

accuracy and robustness of the insulation failure prognosis [61].

103

BIBLIOGRAPHY

104

BIBLIOGRAPHY

[1] DuPont Nomex 410 Technical Datasheet, DuPont, April, 2016.

[2] P. Neti, K. Younsi, and M. R. Shah, “A novel high sensitivity diﬀerential current trans-
former for online health monitoring of industrial motor ground-wall insulation,” in 2013
IEEE Energy Conversion Congress and Exposition, Sep. 2013, pp. 2493–2499.

[3] J. G. Cintron-Rivera, S. N. Foster, and E. G. Strangas, “Mitigation of turn-to-turn
faults in fault tolerant permanent magnet synchronous motors,” IEEE Transactions on
Energy Conversion, vol. 30, no. 2, pp. 465–475, June 2015.

[4] S. Choi, M. S. Haque, M. T. B. Tarek, V. Mulpuri, Y. Duan, S. Das, V. Garg, D. M.
Ionel, M. A. Masrur, B. Mirafzal, and H. A. Toliyat, “Fault diagnosis techniques for
permanent magnet AC machine and drives—a review of current state of the art,” IEEE
Transactions on Transportation Electriﬁcation, vol. 4, no. 2, pp. 444–463, June 2018.

[5] A. E. Ginart, D. Brown, P. W. Kalgren, and M. J. Roemer, “On-line ringing character-
ization as a PHM technique for power drives and electrical machinery,” in 2007 IEEE
Autotestcon, Sep. 2007, pp. 654–659.

[6] J. Pyrhonen, T. Jokinen, and V. Hrabovcova, Design of Rotating Electrical Machines,

2nd ed. John Wiley & Sons, INC, 2014.

[7] G. C. Stone, I. Culbert, E. A. Boulter, and H. Dhirani, Electrical Insulation for Rotating
Machines: Design, Evaluation, Aging, Testing, and Repair, 2nd ed. John Wiley & Sons,
INC, 2014.

[8] V. Sihvo and J. Pyrhonen, “Steam-resistivity of wire insulating materials,” in 2007

IEEE Lausanne Power Tech, July 2007, pp. 34–38.

[9] Y. Li, B. Gao, and J. Unsworth, “Degradation of epoxy insulation under combined high
humidity and elevated temperature,” in Proceedings of 1994 4th International Confer-
ence on Properties and Applications of Dielectric Materials (ICPADM), vol. 2, July
1994, pp. 752–755 vol.2.

[10] A. Griﬀo, I. T. syokhla, and J. Wang, “Lifetime of machines undergoing thermal cycling
stress,” in 2019 IEEE Energy Conversion Congress and Exposition (ECCE), Sep. 2019,
pp. 3831–3836.

105

[11] D. Kavanagh, K. Gyftakis, and M. McCulloch, “Early-onset degradation of thin-ﬁlm
magnet wire insulation for electromechanical energy converters,” in 2019 IEEE 12th
International Symposium on Diagnostics for Electrical Machines, Power Electronics
and Drives (SDEMPED), Aug 2019.

[12] N. Lahoud, M. Q. Nguyen, P. Maussion, D. Malec, and D. Mary, “Lifetime model of
the inverter-fed motors secondary insulation by using a design of experiments,” IEEE
Transactions on Dielectrics and Electrical Insulation, vol. 22, no. 6, pp. 3170–3176,
December 2015.

[13] C. Zoeller, T. M. Wolbank, and M. A. Vogelsberger, “Inverter-fed drive stator insulation
monitoring based on reﬂection phenomena stimulated by voltage step excitation,” in
2016 IEEE Energy Conversion Congress and Exposition (ECCE), Sep. 2016, pp. 1–8.

[14] Inﬁneon. (2015, January) IRFB812PbF HEXFET Power MOSFET. [Online]. Available:
http://www.inﬁneon.com/cms/en/product/power /power-mosfet/20v-300v-n-channel-
power-mosfet/ 120v-300v-n-channel-power-mosfet/IRFB812/

[15] M. Toudji, S. Duchesne, and G. Parent, “Predictive diagnostic based on HF modeling
of electrical machines windings,” in 2019 IEEE International Electric Machines Drives
Conference (IEMDC), May 2019, pp. 901–906.

[16] S. Grubic, J. M. Aller, B. Lu, and T. G. Habetler, “A survey on testing and monitoring
methods for stator insulation systems of low-voltage induction machines focusing on
turn insulation problems,” IEEE Transactions on Industrial Electronics, vol. 55, no. 12,
pp. 4127–4136, Dec 2008.

[17] ——, “A survey of testing and monitoring methods for stator insulation systems in
induction machines,” in 2008 International Conference on Condition Monitoring and
Diagnosis, April 2008, pp. 196–203.

[18] M. Riera-Guasp, J. A. Antonino-Daviu, and G. Capolino, “Advances in electrical ma-
chine, power electronic, and drive condition monitoring and fault detection: State of
the art,” IEEE Transactions on Industrial Electronics, vol. 62, no. 3, pp. 1746–1759,
March 2015.

[19] A. Bhure, E. G. Strangas, J. Agapiou, and R. M. Lesperance, “Partial discharge de-
tection in medium voltage stators using an antenna,” in 2017 IEEE 11th Interna-
tional Symposium on Diagnostics for Electrical Machines, Power Electronics and Drives
(SDEMPED), Aug 2017, pp. 480–485.

106

[20] H. Kim, T. Kong, S. B. Lee, T. . Kang, N. Oh, Y. Kim, S. Park, and G. C. Stone,
“Experience with stator insulation testing and turn/phase insulation failures in the
power generation industry,” in 2017 IEEE 11th International Symposium on Diagnostics
for Electrical Machines, Power Electronics and Drives (SDEMPED), Aug 2017, pp. 21–
30.

[21] P. Zhang, K. Younsi, and P. Neti, “A novel online stator ground-wall insulation monitor-
ing scheme for inverter-fed AC motors,” IEEE Transactions on Industry Applications,
vol. 51, no. 3, pp. 2201–2207, May 2015.

[22] J. Yang, J. Cho, S. B. Lee, and J. Yoo, “An advanced stator winding insulation quality
assessment technique for inverter-fed machines,” in Conference Record of the 2006 IEEE
Industry Applications Conference Forty-First IAS Annual Meeting, vol. 3, Oct 2006, pp.
1562–1569.

[23] V. P. Metelkov, M. Y. Borodin, K. A. Kondakov, and K. E. Nesterov, “Method for
monitoring the condition of the motor winding insulation,” in 2018 17th International
Ural Conference on AC Electric Drives (ACED), March 2018, pp. 1–4.

[24] Sang Bin Lee, Jinkyu Yang, K. Younsi, and R. M. Bharadwaj, “An online ground-
wall and phase-to-phase insulation quality assessment technique for AC-machine stator
windings,” IEEE Transactions on Industry Applications, vol. 42, no. 4, pp. 946–957,
July 2006.

[25] I. Tsyokhla, A. Griﬀo, and J. B. Wang, “On-line condition monitoring for diagnosis
and prognosis of insulation degradation of inverter-fed machines,” IEEE Transactions
on Industrial Electronics, pp. 1–1, 2018.

[26] P. Nussbaumer, M. A. Vogelsberger, and T. M. Wolbank, “Induction machine insula-
tion health state monitoring based on online switching transient exploitation,” IEEE
Transactions on Industrial Electronics, vol. 62, no. 3, pp. 1835–1845, March 2015.

[27] C. Zoeller, M. A. Vogelsberger, and T. M. Wolbank, “Monitoring of thermal degraded
AC machine winding insulation by inverter pulse excitation,” in 2016 XXII International
Conference on Electrical Machines (ICEM), Sep. 2016, pp. 1119–1125.

[28] I. Tsyokhla, A. Griﬀo, and J. Wang, “On-line motor insulation capacitance monitoring
using low-cost sensors,” in 2019 IEEE Energy Conversion Congress and Exposition
(ECCE), Sep. 2019, pp. 6996–7003.

[29] J. Cintron-Rivera, “Fault mitigation in permanent magnet synchronous motors
copyright

turn-to-turn failure,” Ph.D. dissertation,

under an internal

2014,

107

in
[Online]. Avail-
http://ezproxy.msu.edu.proxy1.cl.msu.edu/login?url=https://search-proquest-

- Database copyright ProQuest LLC; ProQuest does not claim copyright
the individual underlying works; Last updated - 2019-10-19.
able:
com.proxy1.cl.msu.edu/docview/1527595438?accountid=12598

[30] M. Sumislawska, K. N. Gyftakis, D. F. Kavanagh, M. D. McCulloch, K. J. Burnham, and
D. A. Howey, “The impact of thermal degradation on properties of electrical machine
winding insulation material,” IEEE Transactions on Industry Applications, vol. 52,
no. 4, pp. 2951–2960, July 2016.

[31] M. Hologne, P. Bevilacqua, B. Allard, G. Clerc, H. Morel, H. Razik, A. Barri`ere, V. Kar-
ode, and N. Devadass, “An experimental approach to the health-monitoring of a silicon
carbide MOSFET-based power module,” in 2017 IEEE International Electric Machines
and Drives Conference (IEMDC), May 2017, pp. 1–7.

[32] J. Liu, G. Zhang, Q. Chen, L. Qi, Y. Geng, and J. Wang, “In situ condition monitoring of
IGBTs based on the miller plateau duration,” IEEE Transactions on Power Electronics,
vol. 34, no. 1, pp. 769–782, Jan 2019.

[33] H. Luo, F. Iannuzzo, and M. Turnaturi, “Role of threshold voltage shift in highly
accelerated power cycling tests for SiC MOSFET modules,” IEEE Journal of Emerging
and Selected Topics in Power Electronics, pp. 1–1, 2019.

[34] U. Karki and F. Z. Peng, “Precursors of gate-oxide degradation in silicon carbide MOS-
FETs,” in 2018 IEEE Energy Conversion Congress and Exposition (ECCE), Sep. 2018,
pp. 857–861.

[35] N. Stojadinovic,

I. Manic, V. Davidovic, D. Dankovic, S. Djoric-Veljkovic,
S. Golubovic, and S. Dimitrijev, “Eﬀects of electrical stressing in power vdmosfets,”
Microelectronics Reliability, vol. 45, no. 1, pp. 115 – 122, 2005. [Online]. Available:
http://www.sciencedirect.com/science/article/pii/S0026271404003786

[36] U. Karki and F. Z. Peng, “Eﬀect of gate-oxide degradation on electrical parameters
of power MOSFETs,” IEEE Transactions on Power Electronics, vol. 33, no. 12, pp.
10 764–10 773, Dec 2018.

[37] T. Azoui, P. Tounsi, P. Dupuy, J. M. Dorkel, and D. Martineau, “Numerical and exper-
imental results correlation during power MOSFET ageing,” in 2012 13th International
Thermal, Mechanical and Multi-Physics Simulation and Experiments in Microelectron-
ics and Microsystems, April 2012, pp. 1/4–4/4.

108

[38] V. Mulpuri and S. Choi, “Degradation of SiC MOSFETs with gate oxide breakdown
under short circuit and high temperature operation,” in 2017 IEEE Energy Conversion
Congress and Exposition (ECCE), Oct 2017, pp. 2527–2532.

[39] X. Ye, C. Chen, Y. Wang, G. Zhai, and G. J. Vachtsevanos, “Online condition monitor-
ing of power MOSFET gate oxide degradation based on Miller platform voltage,” IEEE
Transactions on Power Electronics, vol. 32, no. 6, pp. 4776–4784, June 2017.

[40] Z. Ni, Y. Li, X. Lyu, O. P. Yadav, and D. Cao, “Miller plateau as an indicator of SiC
MOSFET gate oxide degradation,” in 2018 IEEE Applied Power Electronics Conference
and Exposition (APEC), March 2018, pp. 1280–1287.

[41] F. Erturk, E. Ugur, J. Olson, and B. Akin, “Real-time aging detection of SiC MOS-
FETs,” IEEE Transactions on Industry Applications, vol. 55, no. 1, pp. 600–609, Jan
2019.

[42] S. Pu, F. Yang, E. Ugur, C. Xu, and B. Akin, “SiC MOSFET aging detection based on
miller plateau voltage sensing,” in 2019 IEEE Transportation Electriﬁcation Conference
and Expo (ITEC), June 2019, pp. 1–6.

[43] S. Pu, E. Ugur, F. Yang, G. Wang, and B. Akin, “In-situ degradation monitoring of SiC
MOSFET based on switching transient measurement,” IEEE Transactions on Industrial
Electronics, pp. 1–1, 2019.

[44] Vishay.

(2011, March)

IRF520,SiHF520 Power MOSFET.

[Online]. Available:

http://www.vishay.com/docs/91017/91017.pdf

[45] CREE.

(2017,

March)

C3M0280090D

Silicon

Carbide

MOSFET
[Online].
https://www.wolfspeed.com/media/downloads/825/C3M0280090D.pdf

Technology.

C3M

MOSFET

Power
Available:

[46] A. S. Babel and E. G. Strangas, “Condition-based monitoring and prognostic health
management of electric machine stator winding insulation,” in 2014 International Con-
ference on Electrical Machines (ICEM), Sep. 2014, pp. 1855–1861.

[47] F. Schemmel, K. Bauer, and M. Kaufhold, “Reliability and statistical lifetime- prog-
nosis of motor winding insulation in low-voltage power drive systems,” IEEE Electrical
Insulation Magazine, vol. 25, no. 4, pp. 6–13, July 2009.

[48] J. R. Celaya, A. Saxena, C. S. Kulkarni, S. Saha, and K. Goebel, “Prognostics approach
for power mosfet under thermal-stress aging,” in 2012 Proceedings Annual Reliability
and Maintainability Symposium, Jan 2012, pp. 1–6.

109

[49] R. K. Singleton, E. G. Strangas, and S. Aviyente, “Extended Kalman ﬁltering for
remaining-useful-life estimation of bearings,” IEEE Transactions on Industrial Elec-
tronics, vol. 62, no. 3, pp. 1781–1790, March 2015.

[50] G. L. Plett, “Extended kalman ﬁltering for battery management

systems of
LiPB-based HEV battery packs: Part 3. state and parameter estimation,” Journal
of Power Sources, vol. 134, no. 2, pp. 277 – 292, 2004.
[Online]. Available:
http://www.sciencedirect.com/science/article/pii/S0378775304003611

[51] Q. Miao, L. Xie, H. Cui, W. Liang,

“Remaining useful
life prediction of
lithium-ion battery with unscented particle ﬁlter technique,”
Microelectronics Reliability, vol. 53, no. 6, pp. 805 – 810, 2013. [Online]. Available:
http://www.sciencedirect.com/science/article/pii/S0026271412005239

and M. Pecht,

[52] B. Saha, K. Goebel, and J. Christophersen, “Comparison of prognostic algorithms
for estimating remaining useful
life of batteries,” Transactions of the Institute of
Measurement and Control, vol. 31, no. 3-4, pp. 293–308, 2009. [Online]. Available:
https://doi.org/10.1177/0142331208092030

[53] D. A. Pola, H. F. Navarrete, M. E. Orchard, R. S. Rabi´e, M. A. Cerda, B. E. Olivares,
J. F. Silva, P. A. Espinoza, and A. P´erez, “Particle-ﬁltering-based discharge time prog-
nosis for lithium-ion batteries with a statistical characterization of use proﬁles,” IEEE
Transactions on Reliability, vol. 64, no. 2, pp. 710–720, June 2015.

[54] A. El Mejdoubi, H. Chaoui, J. Sabor, and H. Gualous, “Remaining useful life prognosis
of supercapacitors under temperature and voltage aging conditions,” IEEE Transactions
on Industrial Electronics, vol. 65, no. 5, pp. 4357–4367, May 2018.

[55] J. Wang and R. X. Gao, “Multiple model particle ﬁltering for bearing life prognosis,”
in 2013 IEEE Conference on Prognostics and Health Management (PHM), June 2013,
pp. 1–6.

[56] M. Jouin, R. Gouriveau, D. Hissel, M.-C. P´era, and N. Zerhouni, “Particle
ﬁlter-based prognostics: Review, discussion and perspectives,” Mechanical Systems
and Signal Processing,
[Online]. Available:
http://www.sciencedirect.com/science/article/pii/S088832701500504X

vol. 72-73, pp. 2 – 31,

2016.

[57] Y. Hu, P. Baraldi, F. D. Maio, and E. Zio, “A particle ﬁltering and kernel
smoothing-based approach for new design component prognostics,” Reliability
Engineering and System Safety, vol. 134, pp. 19 – 31, 2015.
[Online]. Available:
http://www.sciencedirect.com/science/article/pii/S0951832014002439

110

[58] J. Devore, Probability and Statistics for Engineering and the Sciences. Belmont, CA:

Brooks/Cole Cengage Learning, 2009.

[59] D. Montgomery, Design and Analysis of Experiments. Hoboken, NJ: John Wiley &

Sons Inc., 2013.

[60] Yiu-Wing Leung, “Maximum likelihood voting for fault-tolerant software with ﬁnite
output-space,” IEEE Transactions on Reliability, vol. 44, no. 3, pp. 419–427, Sep. 1995.

[61] J. Sikorska, M. Hodkiewicz, and L. Ma, “Prognostic modelling options

remaining useful
Processing,
vol.
http://www.sciencedirect.com/science/article/pii/S0888327010004218

for
life estimation by industry,” Mechanical Systems and Signal
25,
[Online]. Available:

pp.

1803

no.

5,

–

1836,

2011.

111