ENABLING UNDERWATER BIOLOGICAL CONTACT SENSING SYSTEMS FOR SEA
LAMPREY DETECTION THROUGH CARBON-BASED INTERDIGITATED
ELECTRODES

By

Ian Gonzalez-Afanador

A THESIS

Submitted to
Michigan State University
in partial fulfillment of the requirements
for the degree of

Electrical and Computer Engineering—Doctor of Philosophy

2025

ABSTRACT

Accurate monitoring of sea lamprey populations is critical to enabling the deployment of more

targeted and effective control measures to minimize the impact associated with this species.

This dissertation demonstrates the development of an automated sea lamprey detection

system and the effects of algal-based biofouling on its voltage response.

The system is built around a sensor composed of two exposed carbon-based planar in-

terdigitated electrodes (IDE) functioning as an underwater biological contact sensor. A

microcontroller-based DC measurement system for the detection of lamprey attachment

underwater is presented; measuring voltage instead of impedance reduces cost and signal

processing complexity, making the device more attractive for field deployment. The system

is calibrated to a baseline output voltage, and deviations from this baseline occur when ob-

jects touch the IDE. Validation was done through testing on live adult sea lampreys using

video recordings to correlate lamprey attachments to the sensor response. Three response

types were identified corresponding to different attachments: sustained, short, and sliding-

sustained (video demonstrations are included in this work). The response to sustained

and sliding-sustained attachments showed a characteristic exponential decay, whereas the

response due to short attachments was indistinguishable from measurement noise. Sea lam-

prey size was found to have a weak linear correlation with both response parameters, positive

for the voltage drop and negative for the time constant of the voltage drop. A representative

circuit for the lamprey-sensor interaction is proposed and simulated using element values

calculated from the response parameters. The response of the model shows agreement with

experimental data.

Characterization of the IDE sensor’s response to sea lamprey attachment allowed for

the development of detection algorithms, to automate the process of detecting sea lamprey

from the sensor output. Inherent limitations on the computing power of the microcontroller

unit used to measure the sensor motivated the exploration of low-complexity models for

the task: single-layer artificial neural networks, logistic regression, Gaussian Na¨ıve-Bayes,

decision trees, random forest, and Scalable, Efficient, and Fast classifieR (SEFR). Threshold

models tuned using a multi-objective optimization formulation were also considered. Mod-

els were trained/tuned with a data set generated through live animal testing and presented

accuracies between 80-86%. The models were deployed on an Arduino microcontroller plat-

form and compared in classification accuracy, detection performance, time complexity, and

memory size using real-time detection testing. Classification accuracies between 65-75% were

observed during validation. Models demonstrated capture rates of 63-85% for sea lamprey

attachments and average detection delays of 9-36 seconds. A video demonstration of a real-

time validation test is also presented. Sensor robustness, sensitivity, and detection speed

were identified as areas of improvement for the system, so electrodes were updated to more

durable, conductive carbon-based 3D-printed electrodes, and an updated measurement sys-

tem combining a resistive bridge topology and an instrumentation amplifier was developed

to allow for scaling the sensor response.

The effects of biofilm accumulation on the baseline sensor response were also studied

to simulate field conditions. Sensors were exposed to optimal biofilm growth conditions in

a photobioreactor running a culture of Chlorella sorokiniana MSU, a robust algae species

native to the Great Lakes region. Two responses to biofilm formation on the sensors were

observed: (1) a consistent rise in voltage following the increasing form of the exponential

decay function, and (2) a similar rise in baseline but followed by a period of approximately

exponential decay. These responses were generally related to changes in sensor resistance,

with (1) corresponding to a decrease and (2) to an increase. A study relating these changes in

sensor resistance to biomass accumulation was performed, but did not produce a conclusive

relationship. The sensors’ ability to pick up biological contact after fouling was observed to

be inconsistently affected. The implications of these results for sea lamprey detection and

potential ways to address them are discussed. Overall, this research presents the first step

toward an electronic sea lamprey monitoring system that can provide a detailed view of sea

lamprey activity, enhancing control and conservation efforts across its entire range.

Copyright by
IAN GONZALEZ-AFANADOR
2025

To science.

v

ACKNOWLEDGMENTS

I would like to begin these acknowledgements by recognizing the person who put me on the

path to becoming a researcher, Dr. Gerson Beauchamp. Dr. Beauchamp was my research

advisor during my undergraduate career, and through his rigorous approach to mentoring

helped me develop my confidence as a scholar and engineer. It was also through him that

I was able to meet another instrumental figure in my development, my current advisor, Dr.

Nelson Sep´ulveda. It is thanks to Dr. Sep´ulveda’s continuing support and mentorship that

I have been able to make it this far in my academic career. He has provided me with a

nurturing environment in which I’ve been able to grow not only as a scientist but also as a

collaborator and peer mentor. I would also like to thank my friends and Applied Materials

Group alumni Dr. Henry D´Souza, Dr. Jos´e Luis Figueroa-Soto, and Dr. Juan Pastrana

who welcomed me into the group with open arms and made me feel at home at MSU. To

the current AMG group members Gerardo Morales-Torres, Yusuf Bin Yaacob, and Luis

Colon Santiago I extend my thanks as well for the good times spent together, the productive

conversations regarding our research projects, and for trusting me to help guide you through

the beginning of your graduate school journeys like the group alumni did for me.

I would also like to thank the members of my committee for making this document

possible: Dr. Xiaobo Tan, the first faculty member to join my graduate committee, whose

insight has been proven invaluable when facing roadblocks in my research, Dr. Wei Liao,

who inspired me to adapt my sensors for biofilm detection in the first place and who has

supported me in those endeavors, and Dr. Saiprasad Ravishankar, who opened my eyes to

the possible applications of the signal processing field to my research.

There are also people from my personal life who have supported me through this odyssey

whom I would like to acknowledge, beginning with my parents, Angel and Lourdes, who have

always pushed me to be the very best I can be while providing a comforting environment if

and when I did not live up to expectations. My siblings, Angeliz, Chaterly and Yansel, have

also been instrumental in my development, serving as confidants and sources of unexpected

vi

inspiration. My long-time friends Alejandra and Carlos Nahyr, have been a constant source

of emotional support, encouragement, and laughs throughout my time in graduate school.

My close friend Alexandra, who always pushes me to take care of myself and take time to

enjoy life in the moment. Lastly, I would be remiss if I did not mention my cat, Isabel, who

brings much needed joy and tenderness into my life.

To all those mentioned here, who have helped me become who I am today, I extend my

most sincerest thanks.

vii

TABLE OF CONTENTS

CHAPTER 1

INTRODUCTION . . . . . . . . . . . . . . . . . . . . . . . . . . .

CHAPTER 2

BACKGROUND, DESIGN METHODOLOGY AND DEVICE
FABRICATION . . . . . . . . . . . . . . . . . . . . . . . . . . . .

CHAPTER 3

IDE SENSOR SEA LAMPREY CONTACT RESPONSE CHAR-
ACTERIZATION AND MODELING . . . . . . . . . . . . . . . .

CHAPTER 4

REAL-TIME AUTOMATIC SEA LAMPREY DETECTION SYS-
TEM USING MACHINE LEARNING CLASSIFIER MODELS
ON EMBEDDED SYSTEMS . . . . . . . . . . . . . . . . . . . . .

CHAPTER 5

THIRD GENERATION MEASUREMENT SYSTEM AND SEN-
SOR DESIGN . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

CHAPTER 6

CHARACTERIZATION OF SENSOR RESPONSE TO ALGAL
BIOFILM ACCUMULATION . . . . . . . . . . . . . . . . . . . .

CHAPTER 7

SUMMARY . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

BIBLIOGRAPHY . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

APPENDIX . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

1

8

17

35

61

66

83

85

91

viii

CHAPTER 1

INTRODUCTION

This work focuses on the development of automatic sensing systems for detecting and quanti-

fying interactions between aquatic organisms and underwater surfaces. The effort is targeted

towards detecting invasive sea lamprey (Petromyzon marinus), parasitic fish that present

surface-attaching behaviors, in the Great Lakes region.

In addition to its application in

invasive wildlife monitoring, a system capable of monitoring biological interactions with

underwater surfaces has potential applications that also motivate this study.

1.1 Invasive Sea Lamprey Detection

Sea lampreys (Petromyzon marinus) are an invasive fish species in the Laurentian Great

Lakes region. They live as parasites for a large part of their life cycle, feeding on other fish

by attaching to them using their suction cup-like mouth (as seen in Fig.1.1-A) and scraping

a hole through the skin to suck on the prey fish’s blood and other bodily fluids [3] leaving

wounds like those shown in Fig.1.1-B. While it is debated whether sea lampreys are native to

Lake Ontario, they have been reported in the lake since at least 1888 [4]. It is believed they

spread from Lake Ontario to Lake Erie, and later to the rest of the Great Lakes, through the

Figure 1.1 A) Sea Lamprey attaching to view pane of a holding tank at Hammond Bay Bi-
ological Station B) Sea Lamprey wounds on a Chinook salmon by Wikimedia Commons
user Fungus Guy[1], reproduced here without alterations under the Creative Commons
Attribution-Share Alike 3.0 Unported license.[2].

1

Welland Canal which began operation in 1829. The presence of sea lampreys in Lake Erie

was first confirmed in 1921 [5], and by 1938 they had spread to the remaining Great Lakes [6].

They quickly exploded in number and contributed to a decline in large fish stocks causing

a devastating decrease in commercial fishing yields, which resulted in the effective collapse

of this industry in many parts of the region [7]. Lake trout (Salvelinus namaycush), lake

whitefish (Coregonus clupeaformis), and burbot (Lota lota), which were lampreys’ preferred

prey, were depleted or extirpated from much of the Great Lakes [8]. As these were the main

predators in the region, their disappearance opened the door for other invasive species such as

rainbow smelt (Osmerus mordax ) and alewife (Alosa pseudoharengus) which created intense

competition for species not directly threatened by sea lampreys [7], further destabilizing the

ecological balance of the region. Through aggressive control measures, including deploying

physical [9] and electrical barriers [10, 11] to limit access to spawning areas, lamprey-specific

pesticide (lampricide) application [12], and trapping and sterilizing male sea lamprey [13],

sea lamprey numbers have been effectively reduced. Accurately monitoring sea lamprey

populations is critical to adequately address the catastrophic threat they pose to the region

and its fishing industry which employs around 75,000 people and is valued at USD 7 billion

annually [14].

Currently, the sea lamprey control program assesses its effectiveness via an adult lamprey

abundance index generated by the capture of adult sea lamprey from tributaries around the

Great Lakes. Adult lamprey traps in these tributaries are staffed for daily operations from

April through June (sometimes through July) [15]. Because of operation cost and poor ef-

ficacy in many streams trapping is conducted in only twenty-nine tributaries across all the

Great Lakes. Other tools have been considered to improve or expand this assessment effort

but have limitations. For example, statistical models have been trained to forecast the prob-

ability of meeting control targets using abundance index information and historic measures

of sea lamprey control effort [16]. While this provides a useful metric for planning control

strategies, it is still limited by the available data. Environmental DNA (eDNA) approaches,

2

where water samples from the field are analyzed to determine the concentration/presence of

genetic material from the target species, have been demonstrated to be able to detect the

adult and larval sea lamprey in lab and field settings, though it is not possible to distinguish

between the two [17]. Further, while there is evidence that the concentration of genetic ma-

terial is positively correlated with adult sea lamprey population densities in the lab [18], in

practice samples can provide presence or absence data for an entire river but do not provide

location, density, or even timing. Dual-frequency identification sonar (DIDSON) has also

been used to successfully track lamprey entry into a tributary of Lake Huron [19], though the

high cost of the equipment (upwards of 25,000 USD per camera) and specific requirements

for deployment (e.g., water depth, non-turbulent environment) make it suited for limited

deployment only. A low-cost, effective tool to assess the presence or relative abundance of

adult sea lamprey in streams could benefit the lamprey control program assessment efforts

and provide a means to better understand the timing of lamprey migration and distribution

of adult sea lamprey throughout the Great Lakes tributaries.

Remote lamprey sensing systems present an alternative surveillance technique; these

systems could factor in environmental disturbances in their measurements, enhancing opera-

tional capabilities by allowing for continuous population monitoring. Moreover, by exploiting

the behavior of the species during the design of these systems, it results in a less disruptive

assessment technique than tagging as well as reduces dependence on traps. In addition to

workload reduction for conservation agencies like the SLCP, trap-free detection could pro-

vide greater insights into the behavior of these fish at times and locations where capture or

direct observation has been historically limited. However, the lack of commercially available

sensors that can detect lampreys and the necessary scale of deployment for these systems to

be effective are major hurdles in the development of such systems. A review of the state of

the art in electronic lamprey detection is included as part of Chapter 2.

The automated sea lamprey detection system presented in this work is meant to be de-

ployed in rivers and tributaries of the Great Lakes region. Any time a system or structure

3

is submerged in (or used to contain) water, its surfaces can interact with the myriad of

microorganisms that call that water home. This can lead to the undesired accumulation

of biological matter on these surfaces, a process known as biofouling [20], which can cause

structural and operational deficiencies. An example of the biofouling formation process as it

occurs in a marine environment is presented in Fig. 1.2. Biofouling like this will likely be an

issue in the deployment of an automated sea lamprey detection system based on this work,

motivating the study of the response of these sensors to the accumulation of biofilm on their

surfaces. Extending the capabilities of the measurement system to the sensing of accumula-

tion of biological material can expand the possible applications of these sensors as biofouling

is a pervasive issue in many fields including food production [?], medicine [21], industrial

processes [22] and algal cultivation systems [23]. Current approaches to biofilm monitoring

rely on staining techniques for visual confirmation of total biomass film or differentiating

between live and dead cells as well as various assays analyzing cultivation of biofilm samples

[24]. These techniques provide good characterization of biofilms at discrete time instances

but do not allow for online monitoring of the build-up. The operating principle behind the

sea lamprey contact detection sensors could be an alternative to traditional methods, provid-

ing real-time information on the accumulation of biological material on underwater surfaces

that can be used as a feedback signal for prevention and control measures.

Figure 1.2 Typical growth process of marine biofouling. Reproduced with permission from
[25].

4

1.2 Problem description and motivation

Differentiating between contact sources poses a challenge for typical contact sensing mech-

anisms such as mechanical switches and pressure sensors. Comprehensive characterization

of the contact source using purely mechanical or pneumatic sensors is simply not possible,

as the physical principles behind these sensors will result in identical responses regardless

of source. Approaches that do offer some selectivity by monitoring the electrical properties

of the sensor, such as capacitive touch sensing, are typically vulnerable to false readings or

noise in the presence of water. The problem of developing a selective underwater contact

sensing system is addressed in this work through:

• The design of low-cost carbon-based interdigitated electrodes for contact sensing using

relative conductivity differences as a driver for contact source characterization.

• Development of a measurement system taking advantage of these principles and the

characterization of the system’s response to sea lamprey attachment.

• Exploring the correlation between characterized response parameters and sea lamprey

size parameters.

• The refining of the measurement system through signal conditioning and data process-

ing techniques for improved performance.

• The integration of the system with machine learning algorithms for automating the

detection of sea lamprey using the sensing system.

• Determining the effect of biological accumulation on the baseline response of the sensor.

To this end, an iterative design process including fabrication, simulations, experiments,

and analysis was carried out to present solutions to these problems. The following chapters

provide fabrication, coding, and data processing guidelines that can be used to build similar

systems as well as a methodology to incorporate machine learning approaches in similar

applications through embedded systems.

5

1.3 Thesis Statement

The major contribution of this work is the development of an underwater contact sensing

system that can distinguish between underwater biological and non-biological contact sources

and automatically detect sea lamprey attachment on these surfaces, as well as study on the

impact of biofilm accumulation on the response of the system

1.4 Outline

The structure of the document is as follows: Chapter 2 provides a review of the state of

the art in electronic sea lamprey detection as well as the rationale behind the selection of

the IDE structure for the task. This is followed by a discussion of the design methodology

for the sensors including the operating principles behind the IDE sensor, material selection

for the device and the differences with traditional IDE applications. The chapter ends with

an overview of the fabrication process for the first two iterations of interdigitated electrode

sensors presented in this work Chapter 3 presents the characterization of the response of

the first-generation sensor design to sea lamprey attachment. Modeling of the response is

performed and is used to undertake a correlation study investigating the relationship between

sea lamprey size parameters and the response parameters. The result of the correlation

study are discussed and a representative circuit for the sea lamprey sensor interaction is

proposed and simulated. Chapter 4 covers the use of learning models deployed on embedded

systems to automate the detection of the sea lamprey based on the characterized response

and contains a discussion on the model selection, how they were trained and deployed as

well as the evaluation of their performance during a real-time sea lamprey detection test.

Chapter 5 briefly presents the third-generation sensing system design, including an updated

measurement approach to allow for amplification of the sensor response. It also includes the

design and fabrication of the final iteration of the IDE sensor based on conductive 3D-printed

electrodes. Chapter 6 covers the attempted characterization of the response of the sensing

system to algal biofilm accumulation to better understand how the system will perform in

the field. It also includes a discussion on the implications of these results for the deployment

6

of the sensors in the field and some suggestions for future research directions. The document

ends with a summary of the contributions and main takeaways of this work.

1

1Some content in chapters 2-4 is adapted from modified versions of Invasive Sea Lamprey Detection and
Characterization Using Interdigitated Electrode (IDE) Contact Sensor - IEEE SENSORS (2021)[26] © 2011
IEEE; Real-Time Invasive Sea Lamprey Detection Using Machine Learning Classifier Models on Embedded
Systems - Neural Computing and Applications (2024)[27] © 2024 Springer Nature and is reproduced here
with the permission of the copyright holders. In reference to IEEE copyrighted material which is used with
permission in this thesis, the IEEE does not endorse any of Michigan State Universities’s products or services.
Internal or personal use of this material is permitted.

7

CHAPTER 2

BACKGROUND, DESIGN METHODOLOGY AND DEVICE FABRICATION

This chapter provides a review of the current state of the art in electronic sea lamprey

detection, as well as a review of interdigitated electrode sensors and their applications. It also

features the design methodology used to develop the contact sensors, including the working

principle, material considerations, and design parameters. Finally, the sensor fabrication

process is presented for both generations of sensors.

2.1 Background

2.1.1 Electronic Sea Lamprey Detection Sensors

Recent advances in electronic sensors for sea lamprey detection [28, 29, 30, 26] show potential

for use in wide-scale electronic monitoring systems that could bypass some of the highlighted

challenges in the expansion of assessment capabilities. These sensors look to exploit sea

lamprey behaviors to enable the use of transducers to detect their presence. As sea lampreys

commonly attach to structures using their suction cup-like mouths, pressure and contact

Figure 2.1 . A) Schematic illustration of the soft capacitive pressure sensor array with air
gap channels and conductive polymer PEDOT:PSS electrodes. B)Photograph of a 12 × 12
capacitive pressure sensor array. Scale bar: 1 cm. reproduced with permission from [28] ©
2019 John Wiley & Sons.

8

Figure 2.2 A) Schematic of a 4-by-4 piezoresistive pressure sensing matrix. B) Sea lamprey
interacting with the sensing array. Reproduced with permission from [29] © 2022 IEEE.

sensing are natural avenues to explore for their detection.

Flexible pressure sensing arrays are a promising candidate for sea lamprey attachment

detection thanks to their ability to generate profiles of the pressures exerted at their surface

[31], from which the suction action of the lampreys can be identified. Several transduction

mechanisms can be used to create these sensors, including capacitance, piezoresistivity, and

piezoelectricity. Shi et al.

[28] reported a soft capacitive pressure sensing array, shown in

Fig.2.1 A, capable of detecting positive and negative pressures which could potentially be

used for sea lamprey detection. The device was manufactured almost completely from flexible

polymers (as seen in Fig.2.1B) and laboratory testing of these sensors using a polymer suction

cup produced promising results; however, the quickly changing pressure profile exerted by

the lamprey [32] proved difficult to capture with the implemented measurement system

[33]. Additionally, capacitive interference due to the dielectric nature of biological material

may have also contributed to difficulties with these measurements. These issues are less

prevalent in piezoresistive sensors, which utilize materials that react to changes in pressure

with a corresponding change in resistance as the sensing mechanism instead of relying on

changes in capacitance driven by deformation under pressure. Indeed flexible encapsulated

piezoresistive array structures, whose construction is shown in Fig.2.2A, reported by Shi et al.

[29] have been shown to be successful at capturing sea lamprey attachments on their surfaces

(as seen in Fig.2.2B). These sensors are not without their drawbacks, as they suffer from

9

crosstalk and hysteresis issues that complicate the processing of their output. Additionally,

the manufacture of these devices can be a time-consuming process, tough advances have

been made in streamlining their fabrication [34]. Finally, flexible encapsulated polypropylene

ferroelectret (PPFE) devices, which present a quasi-piezoelectric effect due to charged voids

in the film as seen in Fig 2.3, have also been shown to be capable of measuring positive

and negative pressures underwater[30] and could be deployed in array structures to generate

similar pressure profiles. The output of these types of sensors could be analyzed using a

computer vision algorithm to detect and characterize the profile generated by sea lamprey

attachment [35]. While promising, the use of computer vision as the underlying detection

algorithm does limit the system due to the computational resources these algorithms require.

Figure 2.3 SEM image of the cross-sectional view of a PP-based ferroelectret device. The
inset shows a PPFE sensor. Reproduced with permission from [30] © 2022 IEEE.

2.1.2

Interdigitated Electrodes as Electrochemical and Biosensors

The development of the sensing system presented in this work took inspiration from the

interdigitated electrode (IDE) electrode structure, which is widely employed in electrochem-

ical sensors. The working principle behind these devices uses the change in impedance seen

across the electrodes, caused by the surface accumulation of the compounds, as the sensing

10

signal [36]. These sensors have been widely utilized as chemical detectors [37] and biosensors

[38, 39] using a technique known as electrochemical impedance spectroscopy (EIS) analysis

to perform label-free detection and concentration measurements. As the surface accumu-

lation that drives the sensing signal could be described as the aggregate of multitudes of

incredibly small discrete “contact” events [40], it is possible this sensing principle could be

adapted on a larger scale for the task of underwater biological contact sensing.

2.2 Design Methodology

2.2.1 Operational principle of IDE Contact Sensing System

Figure 2.4 Visual representation of IDE contact sensor operating principle.

One of the main contributions of this work is presenting a novel use of the IDE structure

as an underwater contact sensor for invasive sea lamprey detection. Figure 2.4 provides a

visual showing the operating principle used to enable contact sensing underwater with the

IDE devices. The sensors are composed of two separate planar electrodes with interlocking

fingers in close proximity but electrically isolated when in air. They are encapsulated in a

way that only the sensing electrode area is exposed to water. To measure the sensors, a direct

current (DC) probing voltage is applied to the sensor and a reference resistor connected in

series, forming a voltage divider circuit. The voltage across the electrodes is monitored and

any abrupt changes to electrode impedance caused by attachment/detachment of the sea

lamprey will be reflected in this output. This approach was selected over more sophisticated

11

impedance measurement schemes, which require additional components and signal processing

because the precision requirements for simple contact detection are lower than those for

electrochemical characterization and detection. Furthermore, the additional components

would increase the cost and complexity of the device making it less suitable for the large-scale

deployments necessary for the system to be effective at monitoring sea lamprey populations.

2.2.2 Differences from traditional IDE applications

While electrochemical IDE sensors use an alternating current (AC) probing signal to directly

determine the impedance/capacitance (or its frequency response) during EIS [41], the ap-

proach taken here is akin to an indirect impedance measurement; since the probing signal

is DC, the steady state value of the response is assumed to be due to the effective resis-

tance between the electrodes with any capacitive behavior being captured in the transient

response of the sensor. When the sensor is submerged in water, an electrical connection is

formed between both electrodes resulting in a measurable baseline, which can be disturbed

during interaction with a foreign object, e.g., lamprey contact. This mechanism introduces

a degree of selectivity in the contact sensing, as the system will react differently to contact

sources with different conductivities, relative to the conductivity of the surrounding medium

—in this case, lake water. This behavior opens the door to characterization of the contact

source which is important as the American brook lamprey (Lampetra appendix) [42], north-

ern brook lamprey (Ichthyomyzon fossor)[43], chestnut lamprey (Ichthyomyzon castaneus)

[44] and the silver lamprey (Ichthyomyzon unicuspis) [45] are all native to the region and

share behavioral patterns with sea lamprey that may lead to their interaction with these

sensors. Electrochemical IDE sensors enforce selectivity by functionalizing the electrode

surface to have an affinity for bonding with the target molecule [46].

In the case of the

device presented here, selectivity is partially enforced simply by the detection mechanism, as

non-lamprey species are not likely to interact with the sensor for a sustained time. During

the adult life stage targeted by this work, the major distinction between native and invasive

species of lamprey is their size, with adult sea lamprey typically being twice as large (35 – 60

12

cm) as the native species (10.3 – 32.6 cm) [47]. The need to determine if there is a correlation

between the size of the lamprey and the response parameters motivates the correlation study

in chapter3.

2.2.3 Sensor Materials and Design Process

Figure 2.5 A) Schematic of first-generation IDE sensor geometry and dimensions. B) Ex-
ample of a manufactured device with a first-generation design. C) Schematic of second-
generation IDE sensor geometry and dimensions. B) Example of a manufactured device
with a second-generation design. Reproduced with permission from [26] © 2021 IEEE.

Clear acrylic panels were used as the substrate for the IDE sensor to allow visual monitor-

ing of lamprey attachment events and correlate these to the electrical measurements of the

system. The specified operating conditions of the sensor preclude utilizing any conductive

materials that are prone to corrosion, as this would not only hamper its functionality but

also lead to unintended contamination of the freshwater systems where it is to be deployed.

Marine-grade conductive carbon spray paint (838AR - Total Ground Carbon Conductive

13

Coating by MG Chemicals) was selected as the electrode material for our prototypes due

to its low cost, relatively high conductivity, and ease of application. The electrodes were

patterned onto the acrylic substrates using 0.6 mm thick 3D-printed stencils developed in

AutoDesk’s Solidworks 3D modeling suite. This fabrication process allowed for an itera-

tive approach at the device design stage, where we tested different geometries, overall size,

electrode finger length and width, and distance between the electrical lead attachments and

noted the effect on the magnitude of sensor response to contact from a conductive probe and

a human hand. Multiple prototype generations were tested until reaching the final design

which showed the best response magnitude under test conditions.

A schematic of the first iteration of the IDE device, which was used in the characteriza-

tion study presented in 3, is shown in Fig.2.5A. The design consists of a small rectangular

interdigitated electrode (sensing area of 32 mm x 49.50 mm) with wingtips to allow for at-

tachment of electrical leads away from the sensing area. During preliminary experiments, it

was observed that the sensitivity of the sensor increased with decreasing distance between

the electrical leads (see Fig. 1-E); and the sensor design/configuration used (Fig. 1-A) al-

lowed for the smallest possible distance that will still allow enough sensing area to include

the lamprey’s oral disk. Wingtips were added to the design to facilitate the protection of the

sensor-lead connection without having to sacrifice the sensing area. It was found that a gap

of 2 mm between electrodes would facilitate stencil fabrication and any post-processing of the

deposited traces, while also providing a detectable change in resistance upon the lamprey’s

attachment. Trace width was selected to optimize the fill factor of the device under the area

and separation constraints. Fig.2.5B shows an example of a final manufactured device used

from the first generation. After the characterization study, the design was updated into a

smaller design, shown in Fig.2.5C, to further optimize the response magnitude of the devices.

Fig.2.5D shows a manufactured device example from the second generation. This iteration

was utilized when the system was integrated with learning models for automatic detection

as presented in 4.

14

2.3 Sensor Fabrication Process

Figure 2.6 A) Device manufacturing process flow B) Encapsulation and lead protection for
first-generation sensors. C) Encapsulation and lead protection for second generation sensors.
A and B reproduced with permission from [26] © 2021 IEEE.

Fig.2.6-A shows the device fabrication process beginning with a square acrylic panel sub-

15

strate (approximately 102 mm x 102 mm), on which the 3D-printed stencil was attached

using repositionable craft adhesive. A coat of conductive carbon paint was then deposited,

allowed to cure for at least one hour, and followed by a second coat which was cured for at

least 12 hours. By following the manufacturer’s instructions each coat should be approxi-

mately 25 µm thick, giving the device a total thickness of approximately 50 µm. After the

paint was completely cured, the stencils were separated from the acrylic, and any excess

paint between the traces was removed using a cotton swab applicator with small amounts

of isopropyl alcohol. Holes were then drilled just outside of the wingtips, which served as

feedthrough vias for the electrical leads. A small piece of copper tape was soldered onto the

leads and used to create a conductive contact between the carbon trace and the leads. Fig.

2.6B shows how this connection was secured for the first generation of sensors, first by fixing

it to the acrylic plate using Kapton tape, followed by encapsulation using polydimethylsilox-

ane (PDMS) for waterproofing, finishing with strapping and use of high-strength tape for a

robust attachment to the acrylic plate. To fix the position of the leads on the backside of

the plate, small amounts of hot glue were utilized. This encapsulation process was improved

for the second generation, as seen in Fig. 2.6 C, where liquid electrical tape was used to

protect the lead connection, and marine grade silicone sealant was used to fix the leads on

the bottom instead. Finally, two velcro tabs were attached to the back of the acrylic panel,

so that it could be attached to a larger panel housing the measurement system for testing.

After the devices were manufactured, the sensor response to open and short-circuit tests

was confirmed in air using a handheld multimeter. For the open-circuit test, an infinite resis-

tance across the two separate IDEs was measured. For the short circuit test, both electrodes

were connected using a piece of copper tape; and the test confirmed a very low resistance

( 100 Ohms) between both copper tape leads. After validating the device’s operation ex-situ,

their integration in a system for underwater deployment and testing followed which will be

covered in the next chapter.

16

CHAPTER 3

IDE SENSOR SEA LAMPREY CONTACT RESPONSE
CHARACTERIZATION AND MODELING

This chapter covers the characterization of the response to sea lamprey attachment for

the first generation of the underwater IDE contact sensor. This includes a description of

the measurement system and an overview of the testing methodologies used to validate

the device through experiments with live sea lampreys. Data from the experiments were

analyzed to characterize the disturbance caused by lamprey attachment on the sensor and

observe whether it was distinguishable from measurement noise. The experimental results

are presented along with a correlation study between the identified response parameters and

lamprey size parameters to enhance the utility of the sensor by exploring the characterization

of lamprey size from the sensor response. The results of a FEM simulation of the sea lamprey-

sensor interaction are used to confirm the experimental sensor behavior. Finally, a simple

representative circuit modeling the lamprey-sensor interaction is proposed and simulated

using component values extracted from the sensor response data.

3.1 Sensor Measurement Approach

The schematic of the system used to measure the sensors is presented in Fig.3.1-A. A DC

voltage divider circuit was used to convert the change in impedance across the electrodes

to a change in voltage which can be more easily measured. An Arduino microcontroller

platform (highlighted in yellow in Fig.3.1-A and B) was utilized to provide the 5 V DC bias

(VDC) to the voltage divider circuit and measure the output voltage at a 10 Hz sampling

frequency using its integrated analog-to-digital converter (ADC) module. This configuration

also allows for the calibration of the measurement system in the presence of water through

the reference resistor (Rref ). To this end, the device is submerged in water and the reference

resistor is adjusted until the output voltage of the circuit is approximately half of the DC

bias, around 2.50 V. This results in an approximate value of the baseline resistance between

the electrodes (RIDE) due to water and enables a relationship between the output voltage and

17

Figure 3.1 A) Measurement system schematic. B) Implemented measurement system at-
tached to an acrylic panel. Reproduced with permission from [26] © 2021 IEEE.

the change in resistance. Using a voltage divider analysis for the circuit shown in Fig.3.1-A,

the output voltage (Vout) can be calculated as:

Vout = VDC ∗

Rref
Rref + RIDE

(3.1)

This equation only holds at all time instances if the circuit is purely resistive, and since

water and animal flesh are dielectric materials, it is reasonable to expect some capacitive

behavior from the IDE sensor. However, since the probing signal is DC, the capacitance

will come into play only in the transient behavior of the response, meaning that (3.1) will

still be a valid description of the steady state of the signal measured by the MCU unit. It

should be noted that Rref has a fixed resistance value, whereas RIDE has a variable resis-

tance that depends on any contact between the two isolated electrodes. In the experiments

performed, the background RIDE value (i.e., resistance before lamprey’s contact) represents

the water resistance between the electrodes (Rwater). By selecting Rref = Rwater and express-

18

ing RIDE = Rwater + R∆ (where R∆ represents the magnitude of the change in resistance

across the electrodes and its sign indicates whether RIDE is increasing or decreasing) we can

re-write (3.1) as:

Vout = VDC ∗

1
2 ± R∆
Rwater

(3.2)

It should be noted that the addition or subtraction of the R∆
Rwater

term occurs when RIDE

increases or decreases, respectively. Thus, an increase in the resistance across the electrodes

(+R∆) will correspond to a drop in the measured output voltage (Vout ); and, correspondingly,

a decrease in RIDE (−R∆) will increase Vout. This allows for the distinction between two

general types of contact: 1) when R∆ is negative (i.e., RIDE < Rwater, and 2) when R∆

is positive (i.e.,RIDE > Rwater). Further characterization within those two categories is

possible by examining the magnitude of the change in R∆. Fig.3.1-B shows how the system

was implemented for testing. The microcontroller platform and a small circuit proto-board

housing the reference resistors were attached to the top of a larger acrylic panel measuring

304.8 mm x 304.8 mm using velcro tabs for easy replacement/modification if necessary. The

sensor was then attached at the bottom of the acrylic panel, to allow for submerging the

system without exposing the rest of the system to moisture.

3.2 Experimental Procedure

Experiments were performed at U.S. Geological Survey, Great Lakes Science Center’s Ham-

mond Bay Biological Station (HBBS) in Millersburg, MI. HBBS staff provided 20 male

sea lamprey specimens, each tagged with a unique identifier for tracking purposes between

experiments. The length, weight and mouth diameter of the lamprey were measured and

recorded as they were tagged (see Table 3.1). As explained later in Section 3.4, the physical

characteristics of the lampreys shown in this table were used to determine the strength of

correlation between the sensor response and each lamprey size parameter. Such correlation is

of particular interest because it would enhance the monitoring capabilities of the developed

system from simple detection to characterization of an attached lamprey’s size, which may

19

Figure 3.2 A) Sea lamprey holding tanks with sensing panel installed. B) Lamprey attaching
onto sensor during testing; view of the lamprey is obstructed by sensor traces as the image
was taken from the outside of the testing tank. Reproduced with permission from [26] ©
2021 IEEE.

allow distinction between species or life stages.

Experiments were conducted in a 200 L rectangular aquarium tank (Fig.3.2-A) supplied

with aerated water from Lake Huron. The sensing panel was submerged in the tank and

fastened using clamps before tests started. The sensor was calibrated to an output voltage

of approximately 2.5 V, and a baseline measurement was recorded for several minutes before

testing to ensure this baseline did not drift significantly due to water turbulence. The

test procedure consisted of placing a lamprey into the tank and coaxing it (manually) into

attaching to the sensor. A closeup of how the sea lamprey’s mouth looks while attaching to

the sensor during testing can be seen in Fig.3.2-B; as the image was taken from the opposite

side of the sensor, through the tank wall, the lamprey is partially obstructed by the electrodes.

Video and voltage measurements were recorded until the lamprey detached from the sensor.

An audio cue was triggered in the video to synchronize the recording and the data offline. On

the first day of testing, 20 experiments were completed on the sensor, one with each lamprey

of the cohort, with additional baseline measurements taken after halfway through testing.

20

Table 3.1 Sea Lamprey Size Parameters

Tag

Weight (g) Length (cm) Mouth Diameter (mm)

blue-018
blue-009
orange-002
orange-017
white-041
orange-025
white-028
blue-002
orange-004
orange-040
pink-034
pink-027
white-043
pink-035
orange-026
white-037
pink-029
blue-024
pink-028
orange-035

255
246
176
241
200
161
117
158
177
177
144
174
118
143
132
141
124
143
128
87

44
50
42
48
45
39
38
39
40
42
39
45
39
40
40
40
38
40
40
35

44.8
40.6
40.4
39.6
39.6
38.0
37.9
37.9
37.8
37.2
36.9
36.5
36.0
34.7
34.3
34.2
33.9
33.3
33.0
29.3

The baseline measurements were taken again after all tests were completed for validation

purposes. On the second day, 10 additional tests were completed with the lamprey cohort for

validation purposes. Voltage data were then processed in OriginLab using a low-pass filter

with a cut-off frequency of 0.25 Hz to filter out high-frequency measurement noise associated

with ADC readings such as input-referred noise and voltage regulation noise. Finally, a time

offset was applied to the data to synchronize voltage measurements and video recording.

Voltage data and video recordings from these tests were then cross-referenced to determine

which disturbances in the baseline were caused by lamprey attachment events (see links to

videos of the test in Section 11 of the appendix)

3.3 Characterization of Sensor Response to Sea Lamprey Attachment

Three distinct types of responses were identified from the sensor, each corresponding to

a different type of attachment. The following discussion is focused on three tests that

21

Figure 3.3 Sensor response to a sustained sea lamprey attachment event. Reproduced with
permission from [26] © 2021 IEEE.

are representative of these responses, but the remaining experimental results are included

in Section 11 of the appendix. Fig.3.3 shows a representative measurement of the sensor

response to a sustained lamprey attachment. It can be noticed in the video for this test

(link 11 in Section 11 of the appendix), that the attaches to the sensor at t = 37.29 s and

detaches at t = 212.38 s, denoted on the plot through colored markers with green indicating

attachment to the panel and red indicating release. The output voltage of the sensor (Vout)

shows a clear exponential decay to a DC value below the original baseline when the lamprey is

attached, corresponding to an increase in the resistance observed between the IDEs (RIDE),

with a sudden change back to baseline when it detaches. The observed response suggests

that the system is not a purely resistive circuit. Instead, it includes reactive components

that result in a time-dependent response characteristic of a first-order transient circuit, which

is to be expected as water and living tissue are known dielectric materials. This behavior

presents an opportunity to model this type of interaction using discrete circuit components,

which will be discussed in Section 3.4.1.

22

Figure 3.4 Sensor response to short sea lamprey attachment events. Reproduced with per-
mission from [26] © 2021 IEEE.

The next plot shown in Fig.3.4 is representative of the sensor response to brief attach-

ments. During this test, the lamprey was visually observed to be attached to the sensor

multiple times for less than 10 seconds (link 11 in Section 11 of the appendix). Each of

these attachment periods is highlighted on the plot using colored data markers. While there

appear to be transients in the data at these instances, the magnitude of the response of

the sensor is not enough to be clearly distinguished from measurement noise. This observed

behavior suggests that a minimum attachment time is needed for the sensor response to be

detectable, which supports the time dependency of the exponential decay response repre-

sented in Fig.3.3.

Fig.3.5 shows the sensor response when the lamprey slid onto the sensor traces after

attaching partially or completely outside of the sensing area. In the video recording of this

test (link 11 in Section 11 of the appendix), two quick attachments directly onto the sensing

area are observed, before a third attachment occurs where the lamprey only partially covers

the sensor. At this point, the handler slid the lamprey across the device to bring it to a central

23

Figure 3.5 Sensor response to a sliding attachment event. Reproduced with permission from
[26]© 2021 IEEE.

position on the panel. As seen in Fig. 3.4 the quick attachments did not generate a response

distinguishable from measurement noise. However, while the characteristic exponential decay

of a sustained attachment seen in Fig. 3.3 is still observed, it is now preceded by a large

transient spike caused by the lamprey sliding on the sensor. This transient indicates that this

attachment method somehow reduces the resistance seen across the IDE sensor for a very

brief period before it begins to increase – resembling the observation for a direct, sustained

lamprey attachment. The mechanism behind this phenomenon is still not fully understood;

though it may be that the movement of the lamprey is contributing to a quick discharging

of the double-layer capacitance that naturally forms between a conductor and ionic liquid

in contact when subjected to an electric field, briefly creating a current path with resistance

smaller than the ionic solution’s bulk resistance.

3.3.1 Baseline Validation

Since the responses of the sensor to the three attachment types that were identified during

testing have been presented, it is also necessary to examine the baseline measurements of

24

Figure 3.6 Compilation of baseline measurements taken throughout testing. Reproduced
with permission from [26] © 2021 IEEE.

our device to determine its reliability and susceptibility to false positives. The plot shown

in Fig. 3.6 contains the results of each baseline measurement test taken throughout our

experiments. While there exists measurement noise in the baseline, there is a clear and

stable average Vout value for each of the tests. This value does vary slightly from test to

test, indicating some drift, but these could be attributed to variances in water temperature

as the water in the tanks is taken directly from Lake Huron, which experiences some slight

temperature variations throughout the day. Regardless of the cause, these variations not

only appear to be much smaller than the sensor response to a lamprey attachment event,

but they also occur over a timescale of hours meaning they could be easily distinguished

from the sensor’s characteristic exponential decay response. Additionally, this stability of

25

Figure 3.7 A) Simplified geometry used in FEM static and transient simulations B) Electric
circuit schematic for COMSOL transient simulation. Reproduced with permission from [26]
© 2021 IEEE.

the measured baseline was maintained even after a total of 30 experiments across two days,

thus providing strong support to the robustness of the device.

3.3.2 FEM Sea Lamprey-Sensor Interaction Simulation

The exponential decay suggests the lamprey-sensor interaction can be represented by an

RC circuit lumped model. A finite element model (FEM) simulation of the lamprey-sensor

interaction was also performed using COMSOL Multiphysics 5.3a. The 3D geometry used

in the FEM simulation is shown in Fig. 3.7-A.

In order to facilitate meshing and obtain a solvable model, the height of the traces

was scaled by a factor of 20 (from 50 µm to 1 mm) and a pyramid shape followed by a

rectangular prism body was used to represent the lamprey in the simulation. This geometric

simplification was necessary because the considerable disparity in sizes between the smallest

and largest features in the model geometry and the very small regions generated when

conforming a conic lamprey mouth to the traces resulted in an extreme amount of mesh

nodes, making the problem very computationally expensive to run. The rest of the geometry

is to scale with the sensor schematic shown as an inset in Fig. 3.7-A and everything is

26

surrounded by a domain representing water (the back boundaries of this domain are light

blue in the figure). It is important to mention that the length of the lamprey object was

selected as such because this parameter did not impact the resistance results significantly

when increased beyond that value. The electric currents module was used to determine

the resistance between the electrodes of the simulated device under static conditions. A

parameter sweep of the distance between the lamprey and the IDE traces was conducted

to see how the resistance changes as the lamprey gets closer to the sensor. The FEM was

coupled to external circuit elements using the electric circuit module, and Fig. 3.7-B shows

the connection diagram. The electrical parameters used in the simulation, σ the electrical

conductivity and ϵr the relative permittivity, are shown in Table 3.2.

Table 3.2 Electrical properties of materials used in COMSOL simulation

Object

ϵr

σ(S/m)

Lamprey
Carbon Paint
Water
Acrylic
Copper

97.75
4.5
80
3.73
1

2.5*10−3
14.492
2.17*10−2
4* 10−14
5.998* 107

The objective of the static FEM simulation was to examine the conditions under which it

would be possible to recreate the steady state behavior of the sensor, meaning the observed

increase in resistance across the IDE when the lamprey came into contact with the sensor.

The value of conductivity utilized for water is the measured value for the lake water utilized

in the experiments. Since there are currently no values reported for the conductivity of

the lamprey, this parameter was modulated until the resistance of the device in the contact

condition was greater than when the lamprey was away from the sensor. For the carbon paint,

the value of the conductivity as reported by the manufacturer was initially utilized but this

resulted in resistance values that were much smaller than what was experimentally measured

through the voltage divider circuit during calibration (43.17 kΩ). Upon measuring the

conductivity of the traces of the device it was found that the actual conductivity was around

27

Figure 3.8 A) Simulated steady-state resistance of the IDE sensor as a function of separation
between lamprey and sensor B) Simulated transient response of the sensor voltage output.
Reproduced with permission from [26] © 2021 IEEE.

10% of the manufacturer specification, so it was adjusted accordingly for the simulation.

The values of resistance obtained from the separation parameter sweep are presented in Fig

3.8.

There is a definite increase in the resistance seen across the terminals of the device as the

lamprey gets closer under the simulated conditions. However, this behavior was observed

only when the conductivity of the lamprey object was less than the conductivity of the water.

Even after utilizing the measured conductivity for the carbon paint, the simulated resistance

was still smaller than the measured value. This discrepancy could be attributed to the

differences in geometry between the simulation and the actual device. Once the conditions

for the static behavior were established, the transient behavior of the simulated device was

examined. It is important to note that the simulated circuit represents a slightly different

situation than what happened during the experiments. The simulation represents the case

where the lamprey is attached to the sensor before the device is powered and stays attached

after the power is switched on. This means the voltage across the reference resistor will not

reach a baseline corresponding to the resistance of the water before decaying further because

of the lamprey attachment. Instead, the decay that will be observed will be from the supply

28

voltage to the steady state voltage level associated with the lamprey attachment. The results

of the transient simulation are presented in Fig. 3.8-B.

An exponential decay of the output voltage, taken across the reference resistor, which

then settles out to a baseline can be observed for the transient simulation of the system. This

behavior is reminiscent of a parallel RC circuit, just as what was observed experimentally,

and strengthens the case for modeling the lamprey sensor interaction as such a circuit.

However, there is a significant difference in the time constant observed for the simulation,

which is in the order of tens of nanoseconds, and the time constants observed during the

experiments, which were in the order of tens of seconds. One factor that contributes to

this discrepancy is the smaller resistance values obtained for the simulation, which will

affect the time constant for an RC circuit, but this difference is not enough to account

for the discrepancy. One important difference between the simplified model and the actual

lamprey-sensor interaction is that the lamprey object in the simulation does not have the

longitudinal cavity which comprises the lamprey’s digestive tract and breathing apparatus.

It may be that this cavity causes the lamprey to behave almost like a cylindrical capacitor

when attached to the sensor, which can have considerably higher capacitance values than

other capacitor geometries. The capacitance in cylindrical capacitors is also affected by

the length of the coaxial cylinders which is neglected in this model for ease of simulation.

However, as the objective of this simulation is to corroborate the general steady-state and

transient behavior observed experimentally rather than perfectly recreate the interaction,

these results were deemed sufficient.

3.4 Response Modeling and Correlation Study

As explained in the previous section, whenever there was a sustained lamprey attachment

on the developed sensor (whether by direct attachment or sliding) an exponential decay

was observed at the output voltage of the sensor (Vout). This type of response can be

characterized by the time constant of the decay along with the steady-state value of the

response. Considering the sea lamprey attachment period, the response voltage can be

29

modeled as:

Vout = Vbaseline − V∆(1 − exp

−t
τ

)

(3.3)

where τ is the time constant of the decay and Vbaseline − V∆ is the value of Vout at steady

state. These parameters can be extracted from an exponential fit to the measurements where

a sustained lamprey attachment was observed and then used to explore their correlation to

sea lamprey characteristics, such as length, weight, and mouth diameter. The exponential

fit tool from OriginLab was used, which uses the following exponential relationship:

y = y0 + A exp R0t

(3.4)

Seventy percent of the first day’s tests presented the characteristic exponential decay

corresponding to the sustained attachment and sliding attachment cases. With a few excep-

tions, detailed in Section 11 of the appendix, these tests were used for the correlation study.

The other remaining tests showed behavior that resembled that of intermittent contact –

i.e., not sustained mouth attachment – to the sensor (see Fig. 3.4). The baseline voltage

(Vbaseline) was obtained on a test-by-test basis, by taking the average of the response before

the exponential decay occurs. The change in voltage (V∆) was calculated by subtracting

the steady-state value (y0) from the exponential fit from the baseline value. Fig. 3.9-A

shows the result of the correlation study between V∆ and the lamprey size parameters. In

order to reduce spread on the x-axis, the size parameters are normalized with respect to the

median value of each parameter of the entire 20 lamprey cohort. value of each parameter

of the entire 20 lamprey cohort. The values of V∆ are plotted along with the corresponding

error bars obtained from the fit. The plots shown in Fig. 4-A show a weak positive linear

correlation between all size parameters and the V∆ of the measured response, indicating that

larger lamprey had a larger response magnitude than smaller lampreys. The next parameter

that was studied is the time constant, which was obtained from the fit equation as τ = −1
R0

seconds. The results of the correlation study between the lamprey size parameters and τ are

30

Figure 3.9 A)) Correlation study between sea lamprey size parameters and voltage delta
response parameter. B) Correlation study between sea lamprey size parameters and time
constant response parameter. Reproduced with permission from [26] © 2021 IEEE.

shown in Fig. 3.9-B. The plots show a weak negative correlation between the time constant

and the lamprey size parameters, indicating that larger lampreys within our sample produce

sensor responses with smaller time constants, implying smaller capacitance, which corre-

sponds to a faster decay to the attachment baseline. Although these are promising results

toward a contact sensor that can distinguish between invasive and native lamprey species,

it is important to recognize that the relatively small sample size (20 male sea lamprey) used

in this study is not sufficient to establish a generalized relationship between the lamprey

size parameters and the response parameters. Additionally, we are not considering if there

exists any cross-correlation between the lamprey size parameters themselves. Because of the

pattern observed between oral disk size and sensor response parameters, partial attachment

cases where the lamprey is not completely covering the sensor could also be problematic;

in these cases, the sensor could respond as if a smaller lamprey was attached allowing for

possible misclassification. However, this effect may be mitigated by using an array of these

sensors, rather than a single unit.

31

3.4.1 Lumped Circuit Model

The simplest circuit structure that would result in the observed electrical behavior is shown

in Fig. 3.10-A. In this circuit, the attachment of the lamprey is represented as a normally-

closed switch (NC Switch in schematic) that opens upon lamprey attachment, forcing the

series connection of a simple parallel RC circuit to the conduction path. When the switch is

triggered (i.e., switch is opened), the capacitor initially behaves as a short circuit, causing the

current to bypass the parallel resistor (RLamp); but after that initial transient behavior, the

capacitor begins to accumulate charge and behaves more like an open circuit, thus sending

more current to the parallel resistor. This has the effect of gradually introducing the resistor

into the current path, which results in a waveform like the one observed during testing. The

effective values of the electric lumped model shown in Fig. 3.10-A can be extracted from the

exponential fit shown in Fig. 3.3. Since a sudden change in Vout is not observed at the instant

the lamprey attaches, it is reasonable to assume that the baseline resistance somehow stays

in the conduction path. Therefore, the steady-state voltage is the result of the following

voltage divider:

Vout = y0 =

VDC ∗ Rref
Rref + Rwater + RLamp

(3.5)

Note that the value of the resistance due to water is different from Rref , as the baseline

output voltage before the lamprey attaches was not exactly 2.50 V. However, Rwater can

be readily obtained by solving for RIDE in (3.1) using the measured values of Vout before

any lamprey activity is initiated. Once the value of RLamp is obtained, it can be used to

determine the corresponding capacitance value, since τ = RLampCLamp .

Simulation of the proposed circuit using the values obtained from this analysis was per-

formed using LTspice XVII and the response of the circuit is shown in Fig. 3.10-B, along

with the experimental measurement. The simulation shows the exponential decay behavior

when the switch opens (corresponding to the lamprey attachment) and a sudden return to

baseline when the switch is closed (corresponding to the lamprey releasing from the panel).

32

Figure 3.10 A) Schematic of the representative electrical circuit. B) Comparison of simulated
circuit output with measured response. Reproduced with permission from [26] © 2021 IEEE.

The simulation, however, does not capture the behavior of the sensor response to a sliding

attachment, which could be viewed as a ramp input to the system and would require a more

complex circuit lumped model.

3.5 Summary

A simple and effective IDE sensor for lamprey attachment detection was demonstrated.

Three different categories of sensor response were identified, corresponding to different at-

tachment types: sustained, short, and sliding-sustained. Whereas the response due to short

attachments was not distinguishable from measurement noise, the other two response cat-

egories showed a characteristic exponential decay shape. The measured voltage at steady

state and the time constant of the response were identified as parameters that may be used

to obtain information about the lamprey’s physical characteristics, which could allow clas-

sification of the attached lamprey, although confirmation would require further testing with

larger sample sizes. The correlation study performed between the response parameters and

the lamprey size parameters showed a weak linear correlation between them. The voltage

change appeared to have a positive correlation with all the lamprey size parameters studied

(length, weight, and mouth diameter), i.e., the bigger lamprey generated a sensor response

with a larger change in voltage, whereas the time constant had a negative correlation with

33

every lamprey size parameter. Expansion of the system into an array of these devices may

help mitigate possible issues that arise in the case of partial attachments, by expanding

the effective sensing area without sacrificing device sensitivity. A representative circuit to

model the lamprey-sensor interaction was proposed and simulated using element values ex-

tracted from the parameters of the measured experimental response. The simulated circuit

matched the behavior observed in the experimental data for the case of sustained lamprey

attachment. The case of sliding-sustained attachment would require a more complex circuit

lumped model that can account for a ramp-type response. Further improvements to the sys-

tem hardware will be necessary, including additional ruggedization of the sensor, expansion

into sensor arrays, waterproofing for supporting electronics, and the addition of data stor-

age/transmission modules. Software enhancement of the system through the use of machine

learning to improve the identification of invasive sea lamprey is also possible. The device

presented in this work is the first step toward a remote sensing system to continuously moni-

tor lamprey populations in the tributaries of the Great Lakes without the need to physically

capture individual animals.

34

CHAPTER 4

REAL-TIME AUTOMATIC SEA LAMPREY DETECTION SYSTEM USING
MACHINE LEARNING CLASSIFIER MODELS ON EMBEDDED SYSTEMS

This chapter includes an overview of the second-generation sensing system including a sig-

nal conditioning circuit and additional hardware modules bringing it closer to a stand-alone

system. The methodology used to build and process the data sets for training machine learn-

ing classifier models to detect the sensor response to sea lamprey attachment is presented.

The validation of these models through live-animal validation testing is shown. Classifica-

tion results on the testing fraction of the training set are presented to show the expected

performances and are compared to the classification results for the validation tests. The

performance of each algorithm at the detection task is quantified and presented along with

the execution time and memory size for each algorithm. Additional details on the methods

used to deploy ANN models to a microcontroller device are included in the appendix.

4.1 Second Generation Sensing System Design

The Arduino Mega platform was chosen as the control unit for the system because of its

memory size (256 KB of flash memory) and number of I/O ports. SD card and real-time

clock (RTC) modules were included in the system to enable independent data collection with

accurately time-stamped data files to move closer to a stand-alone system fit for field de-

ployment. Communication between the peripherals is done with Serial Peripheral Interface

(SPI) for the digital potentiometers and the SD card reader and Inter-Integrated Circuit

(I2C) for the RTC module. A peripheral board with two buttons and an LED served as

the user interface of the system during testing. One button was used to toggle system data

collection on and off, and the other was used during data collection to indicate when a sea

lamprey is attached to the panel, allowing for real-time labeling of the sensor output data.

The LED was used as a visual indicator for the state of data collection. An additional pe-

ripheral board with a 2 x 2 LED array was used during validation testing as visual indicators

for the output of a selection of the detection models. The Arduino board 5 V supply is used

35

Figure 4.1 Complete sensing system schematic diagram showing connections for signal mea-
surement, data communication and power supply.

to provide power to most elements of the sensing system, though the signal conditioning

circuit is powered directly from an external supply. The schematic for the system can be

seen in Fig. 4.1.

4.1.1 Signal Conditioning

In this updated sensing system a digital potentiometer in rheostat configuration is used

as a programable reference resistor in the voltage divider. This allows the system to self-

calibrate to a voltage baseline of around 2.5 Volts (within a tolerance of +/- 25 mV) using

proportional feedback tuning at the start of each test. Though the sensing area of a single

sensor is limited, an array configuration using individual sensors as “pixels” can effectively

increase the sensing area at the cost of additional components and processing. However,

whenever a multiplexed input (1 sample-and-hold circuit used for all channels) analog-to-

digital converter (ADC) is used to measure multiple channels, care must be taken to avoid

“crosstalk” noise, where one channel’s measurement is reflected on the next channel being

measured. In this system, op-amps in a low-pass voltage follower configuration are used to

decouple the voltage dividers from the ADC input, ensuring a small source impedance that

36

Figure 4.2 Circuit schematic showing the measurement and signal conditioning circuitry for a
single sensor pixel. The output of a voltage divider between the sensor and the programmable
reference resistance is passed through a filtering voltage buffer before being sent to the ADC
module for capture.

allows the sample and hold circuit to discharge properly between measurements. The circuit

schematic representing the measurement of a single “pixel” in the sensor array is shown in

Fig. 4.2.

4.1.2 Device Preparation

Several sensors consisting of a single IDE ‘pixel’ were manufactured to be used in experiments

to build a dataset of the sensor response to sea lamprey attachments for this generation of

design. System components were integrated onto a large acrylic panel for testing as shown

in Fig. 4.3A).

4.2 Data Collection, Labeling, and Processing

Large amounts of data measurements (sensor response to lamprey attachment) are neces-

sary to train a model that will accurately determine when a lamprey has attached to the

panel. This data was obtained from live-animal experiments at the U.S. Geological Survey,

37

Great Lakes Science Center’s Hammond Bay Biological Station (HBBS) in Millersburg, MI.

HBBS staff provided 66 lamprey specimens (33 males and 33 females), each tagged with a

unique identifier for tracking purposes between experiments. The length, weight, and mouth

diameter of the lamprey were measured and recorded as they were tagged for future corre-

lation studies between lamprey size parameters and the sensor response parameters as any

relationships found would be useful in differentiating between lamprey species or life stages

of the animals attaching to the panel. Data collection was performed during two trips: in

the first trip (June 2022) 26 sea lamprey (13 male and 13 female) were tested across three

days on a single pixel sensor to begin building a data set; in the second trip (July 2022), 40

sea lamprey (20 male and 20 female) were tested to enhance the dataset.

4.2.1 Experimental Setup for Sensor Response Data Collection and Labeling

Experiments took place in a 204 L rectangular aquarium tank supplied with aerated water

from Lake Huron. The sensing panel was submerged in the tank and fastened using clamps

before testing began. At the beginning of each test, the Arduino was restarted, which

generated a new timestamped file that would hold the data for the test. The system then

proceeded to self-calibrate itself, indicating when it finished through a serial connection

so that the system operator could begin data recording. Once the operator started the

recording, a handler introduced a lamprey to the tank and attempted to coax it to attach to

the sensor. During the test, the operator pressed the second button on the peripheral any

time the sea lamprey attached to the panel, effectively manually labeling the data in real-

time. Once the sea lamprey is attached to the sensor uninterruptedly for at least 30 seconds)

it is coaxed off the sensor and taken out of the tank, ending the test. The tests were video

recorded to corroborate the manual label if needed. A view of a sea lamprey attachment

from a sample video recording during one of the single pixel tests, where the lamprey can be

seen attaching to the sensor through the gaps in the conductive carbon traces, can be seen

in Fig. 4.3D).

38

Figure 4.3 A) Full view of the sensing system with all components. B) Indicator lights
responding to lamprey attachment during validation testing as seen through the glass tank
wall and the IDE sensor traces.

39

4.2.2 Data Processing

Once testing was completed, the data for all tests were compiled into a single set for use

in training the detection algorithms. Normalization of the training data was performed to

equalize the influence of each input variable on the model since the voltage level of the sensor

is about two orders of magnitude larger than the rate of change. Although the sensor response

to lamprey attachment is generally a pure decay and it is assumed that the baseline will hold

over time due to the self-calibration algorithm, response values greater than the baseline are

possible if the lamprey slides on the sensor. Because of this, the voltage was normalized to a

maximum expected value of 3 volts (baseline + 0.5 V buffer zone), to completely capture the

expected values in a range between [0,1]. For the values of the rate of change, the selection of

the normalization factor is not immediately clear from previous research. This is because the

rate of change is expected to stay relatively small most of the time, even during the decay,

but will have very large positive spikes when the response suddenly goes back to baseline

after the sea lamprey detaches. Additionally, an offset will be necessary to fit the values of

the rate, which can be positive or negative, to the range [0,1]. The statistical information for

the rate data was obtained to aid in selecting the normalization factor and the offset. Since

the distribution of the data points looked to follow a normal distribution profile centered

about 0, the normalization factor was selected as 3 times the standard deviation of the set.

This captured 98% of the data between [-1,1] after which the data was mapped to a range

of [0,-1] using a point-slope equation. Values for the average voltage and rate were computed

using a rolling average with a window size of 5 on the normalized data.

4.3 Classifier Models for Detection Algorithms

There is a consistent trend of combining machine learning models with sensors to create

”smart” systems that can react to or infer things from measured data [48]. While many com-

plex state-of-the-art classifier models exist, this application requires a model simple enough

to be deployed on a microcontroller unit (MCU), where memory and power constraints must

be considered, for wide-scale deployment of the sensing system to be economically viable.

40

Implementations of ML algorithms for resource-constrained embedded systems have been ex-

plored and proven successful [49, 50]. Embedded machine learning (EML) applications have

been demonstrated in many fields, from enabling self-powered chemical oxygen demand sen-

sors [51], to performing tasks such as gesture recognition [52], urban sound classification [53]

and injury prediction [54]. This motivates the exploration of ML algorithms for the problem

of sea lamprey attachment detection, which can be reduced to classifying the sensor response

measurements as corresponding to baseline or attachment. This study focused on ML clas-

sifier models that were computationally simple and did not require much memory storage:

single-layer artificial neural networks, logistic regression, naive-Bayes, decision trees, random

forests, and the scalable, efficient, and fast classifieR (SEFR) [49].

A straightforward thresholding approach (i.e., indicating there is a sea lamprey on the

panel if the input falls within a certain region in the input variable space) was also considered

due to its low computational complexity, making it well-suited for deployment in low-resource

systems. However, manual tuning of the four boundary values to achieve the best detection

performance is not practical. This was resolved by posing the tuning as a multiple objective

optimization (MOO) problem with the goal of maximizing accuracy measures and using an

MOO solver to effectively automate the selection of the thresholding algorithm’s parameters.

Previous work on the IDE sensors has shown they present a characteristic decay response

to a sustained sea lamprey attachment. Thus, an algorithm that can recognize these decays

will be able to detect their attachment onto the panel. The sensor response behaves dif-

ferently during baseline than it does during the lamprey attachment-induced decay, both

in terms of the voltage level and rate of change. This means it might be possible to use

a machine learning classifier model on these measurements to identify when the response

has decay characteristics and when it has baseline characteristics, effectively identifying sea

lamprey attachments. This can be visually confirmed by looking at the scatter plot showing

the normalized average voltage levels and normalized average rate of change for all the data

points in the training data shown in Fig. 4.4. The data points corresponding to baseline

41

Figure 4.4 A) (Top left) Distribution of training set data in the input variable space. B)
(Top right) Close-up of the distribution in the highest density region in the input space. C)
(Bottom left) Distribution of data points that correspond to lamprey attachment only. D)
(Bottom right) Distribution of data points corresponding to baseline only.

42

measurements are colored red, while the blue points correspond to sea lamprey attachment.

The clustering of the data points around particular regions is a promising indicator that

classifier models could differentiate between them. The shape of the clusters also reinforces

the potential viability of a simple threshold approach (i.e., indicate there is a sea lamprey on

the panel once the signal decays beyond a specified threshold) suggesting that a rectangular

threshold model with 4 parameters (upper and lower bounds for each variable) could be

sufficient for the task and should be explored as a computationally simple alternative to the

ML models as well.

All of the machine learning models studied in this work were trained on a desktop com-

puter (Processor: Intel(R) Core (TM) i7-10510U CPU @ 1.80GHz, 2304 MHz, 4 Core(s), 8

Logical Processor(s), RAM: 16 GB) using the sci-kit-learn python package [55]. The rationale

behind their selection as well as details on how they were deployed on the micro-controller

system are covered in the following subsections.

4.3.1 Neural Network Models

A single-layer artificial neural network (ANN) classifier model was chosen because of its

straightforward low-resource implementation (in terms of both memory and computation)

for inference in an embedded system when compared to other models such as support vector

machines (SVMs), k-nearest neighbors (kNN) or deep neural networks (DNN). The model

was deployed on the Arduino Mega using a slightly modified version of the open-source C99

library tinn [56]. This library implements a single-layer ANN with a sigmoidal activation

function. It was modified to include a bias for each individual neuron in the layer, which

matches the ANN model implementation in scikit learn. The model parameters are extracted

from the trained networks and stored as float arrays which are loaded into an empty instance

of the model. As this is a binary classification problem, the choice was made to have a single

output neuron meaning the output of the model is the probability the inputs correspond to

a baseline measurement and not a class label directly. If this value falls below 0.5 then that

data is labeled as a lamprey attachment; it is labeled as a baseline measurement otherwise.

43

Four ANNs of increasing hidden layer size (n = 6, 12, 24, 36) were trained and tested to

explore if performance varied with complexity.

4.3.2 Logistic Regression and Naive-Bayes Models

Logistic regression (referred to as LGR throughout the paper) and Gaussian Naive Bayes

(GNB) classifiers were also considered for the classification task due to their ease of imple-

mentation and low computational complexity. LGR is equivalent to reducing the sigmoidal-

based ANN to a single output neuron with two inputs and no hidden layer effectively being

the simplest network possible. The LGR model was deployed by computing the following

equation with the parameters obtained from training:

P =

exp(B + w1x1 + w2x2)
(1 + exp(B + w1x1 + w2x2)

where P is the probability of the input data x1 and x2 (in this case the average voltage and

average rate of change) being classified as baseline, B is the bias for the LGR, and w1 and

w2 are the weights for each variable determined during training.

The GNB model uses a similar approach, under the assumption that the data classes

follow Gaussian distributions, the log of the probability of the input belonging to each class

is computed, and the class with the highest value is selected as the classification. This is

implemented as,

ypred = argmax

y

(cid:16)

P (y)

log

(cid:17)

P (xi|y)

2
(cid:89)

i=1

where P (y) is the class prior (determined during testing) and P (xi|y) is the class-conditional

probability which for the Gaussian distribution is given as:

P (xi|y) =

1
(cid:112)2πσy

2

(cid:16)

exp

−

(xi − µy)2
2σy

2

(cid:17)

where σy and µy are the Gaussian distribution parameters found for each class. This model

was deployed on the microcontroller using the MicroMLGen python package [57] which

44

enables users to port a variety of models from the scikit-learn package into C header files

for easy deployment on such devices. The model file defines the structure, any required

operations, and trained model parameters.

4.3.3 Tree-Based Models

Another class of models considered for the detection task are tree-based algorithms. These

models can be considered as extensions of the thresholding approach. Their simple structure

based on branching conditional statements, called decision trees (DT), is very simple to

implement and can be constrained to reduce the model size. Ensembles of multiple trees,

known as random forest (RF) models, can also be utilized for inference potentially providing

greater accuracy at the cost of a larger model size. The decision tree model generated was

limited to a depth of 10 levels and 20 nodes. The RF model consisted of 10 of these base

units. The header files for these models were also generated using the MicroMLGen package.

4.3.4 SEFR

The SEFR model was specifically developed for binary classification on low-resource systems[49],

making it ideal for the target application. During training, the algorithm finds the parame-

ters of a hyperplane defined by the following matrix equation:

wT x + b = 0

where w is the weight matrix, x is the input vector and b is the bias.

Inference is then

performed by evaluating wT x + b ≤ 0; if the statement holds the input is classified as an

attachment. This model is not included in scikit-learn, but there are wrappers available to

use the Python implementation with the package and it is compatible with MicroMLGen for

deployment on the Arduino.

4.3.5 Threshold Models

The boundary values for the threshold models were generated through a multi-objective

optimization-based tuning procedure using the NSGA-2 [58] genetic optimization algorithm.

A multiple objective optimization (MOO) problem formulation was selected over traditional

45

single-objective optimization tuning approaches to avoid overfitting that may occur from

using a single (potentially class-biased) accuracy metric. By using NSGA-II [58], a genetic

MOO algorithm with excellent performance solving two-objective optimization problems, two

accuracy metrics can be considered during optimization tuning leading to a more balanced

model. The algorithm takes a group of randomly generated solutions contained within

the specified upper and lower bounds for each variable and modifies them over multiple

generations by computing the objective values for each solution in the population, replacing

the worst-performing solutions with slightly altered combinations of the best ones. This

generates an optimal set of solutions (known as a pareto-front) which the user can then select

from depending on the relative importance of the two objectives. This selection process can

be used to further reduce the influence of a biased metric. The precision, recall, and F1

scores for each class in the data set (attachment and baseline) were considered as targets for

the tuning problem. The most effective pairwise combination of these six metrics was not

immediately clear, so all 15 possible combinations were used to generate threshold models

under the same conditions and their performance was examined to determine the best-

performing pair. As the NSGA-II algorithm is structured to solve minimization problems

the accuracy metrics are rewritten as:

Fi = 1 − fmetrici = 1, 2

Where Fi is the i-th objective function to be minimized, and fmetric is any of the six

accuracy metrics. This forces the algorithm to minimize the difference between the metrics

and their ideal value of 1 (i.e., a perfect score). The threshold values were bounded (between

0 and 1.2 for voltage and -2 and 2 for rate) to obtain reasonable values for the generated

models as well as improving the speed of the search. A population size of 100 solutions

was evolved over 100 generations to obtain models for each pairwise combination of metrics.

The performance of the solutions was evaluated each generation by feeding the training set

(same set used to train the ML models) into the generated models and using Scikit-learn’s

46

Figure 4.5 Classification performance comparison between the 15 possible accuracy metric
combinations on the testing fraction of the training set.

classification report function to obtain the relevant accuracy metrics before comparing them

to the target value. The achievement scalarization function (ASF), which uses a weight vector

to specify the relevant importance of each objective [59] and extract a solution matching the

specified trade-off, was used to select a model from the set of solutions generated by the

algorithm. To simply the selection process, equal importance was assigned to each object

by using 0.5,0.5 as the weight vector. The optimization and solution selected was facilitated

by the use of the PyMOO multiple objective optimization package for Python [60]. The

classification performance of the selected models on the testing fraction of the training set

for each combination of accuracy metric targets is presented in Fig. 4.5.

Several trends in performance can be identified among the generated models. Three

models (identified as 2, 3, and 4 in Fig. 4.5) presented very poor classification performance

overall, perhaps due to antagonism between selected tuning parameters or bias toward one

class, with critically low scores for one of the two classes and were discarded from consider-

47

ation for deployment. Eight of the fifteen models (identified as items 5, 6, 7, 10, 11, 12, 13,

and 15 in Fig. 4.5) show almost identical performance across metrics. These present recall

and precision scores around 0.72 respectively for the attachment class indicating that a large

majority of the points corresponding to the class are captured and that most of the data

points identified as attachments correspond to that class. Models 1 and 8 presented the best

recall scores for attachment of the group, though their precision scores are on the lower end

of the group as well showing the tradeoff in performance between the two metrics. The final

models have distinct performance profiles with 9 at the opposite end of the spectrum show-

ing the highest precision for the attachment class and the lowest recall score of the viable

models and 14 presenting a middle point in performance between the two extremes. After

examining the results from the 15 possible models models 1,9 and 14 were the ones selected

for real-time deployment testing to explore the range of performance that the thresholding

models demonstrated in the training set to represent the range of performance observed. The

first of these, referred to as Thresh 1, was tuned using the recall scores of the attachment

and baseline classes, which resulted in a higher capture rate for attachment data points at

the cost of higher rates of false negatives. Thresh 2 was tuned using the precision for the

attachment class and the f1-score for the baseline class, which resulted in a more conservative

model with fewer false positives but also less attachment capture. The final model, Thresh

3, was tuned using the f1-score and recall score for the baseline class and demonstrated a

performance between Thresh 1 and Thresh 2. The boundary values are presented in Table.

4.1.

Table 4.1 Boundary Values for Threshold Models

Model

Lower BoundV oltage Upper BoundV oltage Lower BoundRate Upper BoundRate

Thresh 1
Thresh 2
Thresh 3

0.32421867,
0.51726076
0.38311498

0.82237967
0.79880256
0.81407758

0.23528082
0.32990316,
0.27848292

1.22225504
0.70216792
0.64207279

To implement these models on the Arduino each data point is evaluated against the

48

bounds and if it falls within the outlined region it is labeled as an attachment.

4.3.6 Testing the Models

The inference time for the models on the Arduino was verified before validating their perfor-

mance. The testing fraction data set was saved as a .csv file on an SD card where it could

be read by the microcontroller. The data was input into the model implementations and the

time to inference was recorded and sent over serial to a PC computer along with the data

and the inferred classification.

The performance of the models at identifying sea lamprey attachment was validated at

HBBS using live animal testing and baseline data collection during July 2023. Animal tests

followed the same procedure as those utilized to build the training set. The raw data was

normalized with the same values used for normalizing the training set, though a longer

averaging window was used (15) due to considerable noise observed in the sensor output

during the validation testing. Baseline data collection is a similar procedure, but no sea

lamprey is introduced into the test tank for the duration of the data collection. In total, 105

data recordings comprised of 97 live-animal tests over 80 different lamprey (40 male and 40

female) and 8 baseline collections, were utilized in the validation of the models. Predictions

by the models were recorded alongside the input data and the manual label data to .csv files

to be analyzed in Python using the scikit-learn package.

4.4 Results

4.4.1 Memory Requirements and Inference Speed

The memory requirements of the deployed models are presented in Table 4.2. All models

required under 3.5 Kb of program memory for their implementation on the microcontroller.

Dynamic memory requirements are also low overall, though the implementation of the ANN

stores the model parameters as program variables in the flash memory resulting in higher

memory usage for these models. The model file column of table4.2 shows the size of the C-

header files defining the model structure, required operations, and trained model parameters

(for all but the ANN models). Even the largest of these files, the random forest model, does

49

Model

NN-6
NN-12
NN-24
NN-36
LGR
GNB
DT
RF
SEFR
Thresh

Table 4.2 Model Memory Requirements

Program Memory
(Bytes)

Dynamic Memory
(Bytes)

Model
(Kilobytes)

File

Size

2698
2794
2988
3182
786
1518
934
3454
870
276

186
282
474
666
17
16
16
16
17
17

6
6
6
6
n/a
3
5
39
1
n/a

not exceed 39 Kb, and the LGR and thresholding models did not require them since their

structures were simple enough to be included directly in the program code.

The inference time, how long it takes for the models to give a prediction, is another

measure of the complexity of the models and could be a key limitation at this scale of

computing power. The inference times were measured by having the deployed models perform

classification on a preprocessed data set (which the Arduino reads from a CSV file using the

SD card adapter) and capturing the time it takes for the model to deliver a classification.

Figure 4.6 shows the inference results for all of the models. Most of the models require

less than 1 ms to compute their decision, and even the largest model, NN-36, achieved an

average inference time of just under 9.4 ms. Since the models are relatively compact, and

their combined inference times are considerably less than the 1 second sampling period of

the system, multiple models can be deployed on the same microcontroller.

4.4.2 Classification Performance

4.4.2.1 Testing Fraction of the Training Set

Table.4.3 shows the classification performance obtained on the testing fraction of the training

set for each of the models in the study. The ANN models showed promising performance

across all levels of complexity, capturing almost half of the data corresponding to attach-

ments (recall), with a high percentage of true positives among those identified as attachments

50

(precision) and an overall accuracy of 84%. The LGR model appears to be the most conser-

vative one with the lowest recall for the attachment class and tying for the highest precision

for that class with the Thresh 2 model. The GNB model’s performance is close to that of

the ANNs, though it appears to be slightly more conservative at assigning attachment labels

and thus captures less of the attachment data. The DT and RF forest models’ performance

was also promising, with higher attachment recall scores than the ANNs while having sim-

ilar precision scores. They are also tied with the Thresh 3 model for the highest overall

accuracy score at 86%. The SEFR model performed similarly to the ANN models in its

recall score but it also delivered one of the lowest precision scores for the attachment class.

The thresholding models each showed distinct performance profiles. Thresh 1 was the most

aggressive model overall capturing 79% of attachment class data though it also presented the

lowest precision score for attachment and the lowest accuracy overall at 80%. Thresh 2 was

one of the more conservative models with the second-lowest recall score but it also showed

the highest precision for the attachment class and an overall accuracy of 83%. Thresh 3’s

Figure 4.6 Average inference time comparison.

51

Table 4.3 Classification Performance in the Testing Fraction of the Training Set

Model

Recallatt P recisionatt F 1att Recallbase P recisionbase F 1base Accuracy

NN-6
NN-12
NN-24
NN-36
LGR
GNB
DT
RF
SEFR
Thresh 1
Thresh 2
Thresh 3

0.48
0.49
0.49
0.48
0.32
0.43
0.67
0.64
0.49
0.79
0.41
0.66

0.84
0.83
0.81
0.84
0.93
0.87
0.77
0.80
0.70
0.60
0.93
0.77

0.61
0.61
0.61
0.61
0.48
0.58
0.72
0.71
0.58
0.68
0.57
0.71

0.97
0.96
0.96
0.97
0.99
0.98
0.93
0.94
0.93
0.81
0.99
0.93

0.84
0.84
0.84
0.84
0.80
0.83
0.89
0.88
0.83
0.92
0.82
0.88

0.90
0.90
0.90
0.90
0.89
0.90
0.91
0.91
0.88
0.86
0.90
0.91

0.84
0.84
0.84
0.84
0.81
0.83
0.86
0.86
0.81
0.80
0.83
0.86

Table 4.4 Classification Performance in Real-Time Validation Testing

Model

Recallatt P recisionatt F 1att Recallbase P recisionbase F 1base Accuracy

NN-6
NN-12
NN-24
NN-36
LGR
GNB
DT
RF
SEFR
Thresh 1
Thresh 2
Thresh 3

0.61
0.61
0.65
0.63
0.48
0.79
0.74
0.73
0.63
0.84
0.53
0.76

0.56
0.56
0.55
0.56
0.60
0.52
0.56
0.56
0.55
0.48
0.64
0.56

0.59
0.59
0.60
0.59
0.53
0.63
0.64
0.64
0.59
0.61
0.58
0.64

0.77
0.77
0.74
0.76
0.84
0.64
0.72
0.72
0.75
0.56
0.85
0.71

0.80
0.80
0.81
0.81
0.77
0.86
0.85
0.85
0.81
0.88
0.79
0.86

0.78
0.78
0.78
0.78
0.80
0.74
0.78
0.78
0.78
0.68
0.82
0.77

0.72
0.72
0.71
0.71
0.72
0.69
0.73
0.73
0.71
0.65
0.75
0.72

performance sits between these two extremes, mirroring that of the DT and RF models.

4.4.2.2 Real Time Validation Tests

The performance of the models at identifying sea lamprey attachment was validated at

HBBS using live animal testing and baseline data collection. Animal tests followed the

same procedure as those utilized to build the training set and the raw data was filtered and

normalized using the same parameters used on the training data. Baseline data collection is a

similar procedure, but no sea lamprey is introduced into the test tank for the duration of the

52

data collection. The fast inference times and low memory requirements of the models allowed

for simultaneous deployment of all models on the microcontroller; every time a measurement

is taken and processed, each algorithm, in turn, uses this data to infer the status of the

sensor. A link (11) to video showing how a selection of four models (NN-24, RF, Thresh

2, and LGR) react through the 2 x 2 indicator array during the validation test (as seen

in Fig 4.3C) is included in Section 11 of the appendix. This inference is compared to the

manual label generated during testing to see how effective the model is at classifying the

sensor state. Table. 4.4 provides a comparison of classification performance for all models

during the real-time validation tests performed at HBBS. The relative performance profiles

the models demonstrated during validation testing follow the general patterns that were

highlighted by the testing fraction classification results. The deployed models were generally

successful at correctly identifying the state of the sensor the majority of the time (model

accuracy ranges from 69-75%). All models, with the exception of the LGR model, were

able to capture more than 50% of the data points corresponding to sea lamprey attachment.

The Thresh 2 model had the highest accuracy score among all of the models, followed by

the decision tree and random forest models. The ANN models followed closely in accuracy

across levels of complexity, as they did for the testing fraction.

Overall, when compared to the classification results in testing fraction, the deployed

algorithms fared better at identifying a larger share of the data points corresponding to

attachment, noted by the higher recall scores for that class across the board. However, this

appears to have come at the cost of a higher amount of false positives as shown by the lower

precision scores. There was a similar hit in performance in the recall score for the baseline

class, though the precision score for that class stayed near the values for the testing fraction.

The overall accuracy of the deployed models, though lower than on the testing fraction, is

still considerable.

53

Figure 4.7 Comparison of the detection performance of each model during the real-time
validation testing.

4.4.3 Detection Performance in Real Time Validation Test

Classification performance measures how well the model sorts individual data points as

baseline or attachment. This will give a general idea of the detection performance of the al-

gorithms, however, the task of detection/monitoring is more complicated because it requires

sustained correct inferences across time. An algorithm that can measure the model perfor-

mance at this task was developed by considering the requirements of the detection tasks.

The algorithm counts an attachment as detected if the model output correctly identifies it

54

at any point in the attachment. Any continuous sequence of false positives is considered

as a single false detection and any sequence of false negatives is considered a single false

baseline instance. The algorithm also reports the time it takes for the model to first detect

an attachment (detection delay) and the length of missed attachments as well.

Fig. 4.7 provides a comparison of detection performance in the real-time validation

tests in terms of the attachment captured and the false positives and negatives generated

during the tests. All of the algorithms managed to capture a majority of the attachments

that happened during validation testing with capture rates from 63-85%. However, they

also generated high numbers of false detections as well as false baseline captures. The

models were trained to respond to the decay generated by sea lamprey attachment which

is a fairly slow process. Considerable low-frequency noise from 0.05-0.12 Hz Hz (as seen in

Fig. 4.8a) of unknown origin was present in the raw measurements across all the validation

tests performed and could have contributed to the high amount of false positives observed.

The models could be reacting to the decaying half cycle of these slow periodic oscillations

and incorrectly identifying them as attachments. Although Fig. 4.8b shows that the noise

at that band has been reduced by the data processing, there was a corresponding increase

in even lower frequency noise which is more likely to be picked up by the models. Further

evidence of this is the fact that the models with the lowest capture rates (LGR and Thresh

2) also have the lowest numbers of incorrect detections, potentially indicating a relationship

between the sensitivity to the characteristic decay and the rate of false positives.

Detection time, how quickly a model first detects a sea lamprey attachment, is an im-

portant performance metric to consider for this task as it highlights a key limitation for

certain applications. Fig. 4.9 shows the timing performance for each model throughout the

validation testing. Fig. 4.9a shows the Thresh 1 model as the fastest at identifying sea

lamprey attachment, with the SEFR and GNB models close behind. However, these are also

the three models with the highest amount of false negatives, again highlighting the trade-off

between speed/capture rate and overall accuracy. The models are slow to react to attach-

55

(a)

(b)

Figure 4.8 A) Frequency spectrum (with the constant component removed) of the raw data
collected during validation. B) Frequency spectrum (with the constant component removed)
of the processed input data for the models.

ment overall; the fastest models required several seconds to react to sea lamprey attachment

and the slowest required more than 30 seconds. This is not unexpected as the decay due

to sea lamprey attachment is a slow process, with decay constants in the tens of seconds

[26], so it takes some time before the sensor response shifts significantly from baseline. Also,

because the speed and amplitude of the decay are positively correlated with the amount of

sensing area covered by the sea lamprey, factors such as fish size and location of attachment

can affect how ”detectable” an attachment is and how long it takes to get there.

For these reasons, it is also useful to look at the length of the attachments that are missed

as another limiting factor. Figure 4.9b), shows the average length of missed attachments

during the validation testing for all of the models. For the activation function-based models

(ANNs, LGR and GNB) and most thresholding models (Thresh 1 and Thresh 2), this value

is slightly larger than the average detection delay. This could indicate that the attachments

missed by these models were those where the position or size of the sea lamprey caused a slow

and low-magnitude response that ended before getting to detectable levels. In the remaining

models (DT, RF, SEFR and Thresh 3) the value is lower than the average detection, possibly

indicating that these are more sensitive to these types of attachments.

56

Figure 4.9 A) Comparison of average delay before detection, B) Comparison of average
missed attachment lengths.

4.5 Discussion

There are several ways a sea lamprey detection system could be utilized to help control the

invasive species in the Great Lakes, each with different performance requirements. The least

demanding of these would be sea lamprey presence detection, where the system would only

be used to determine if sea lampreys are in a particular area. In this application, the output

of the detection algorithm would only be scrutinized to determine if any attachments had

occurred within a specified time frame. Possible applications include, deployment in places

where sea lamprey gather but traps are difficult to deploy to enhance monitoring capabilities;

or deployment in streams with seasonal barriers to better inform installation timing. For this

application, the system should have a low false positive rate, to avoid directing resources

where they are not needed, with detection speed being a secondary consideration. If the

output is further analyzed to count the number of attachments in a given period, an activity

index could be generated that would provide more detailed information regarding lamprey

distribution, enabling more efficient resource allocation. This application would also require

a low false negative count to be viable, as false negatives during an attachment could result

57

in it being counted multiple times, which could also lead to improper resource allocation.

All the models in the study presented a relatively large number of false detections (false

positives and negatives), ranging from 688 to 1953 counts as seen in Fig. 4.7, apparently

making them unsuited for the presence/activity index applications. However, the presence

of low-frequency noise in the input data, which exhibits behavior close to the slow decay

and sudden increase behavior the algorithms were trained to pick up on, likely led to ele-

vated levels of false positives and negatives during validation testing. This noise could be

addressed with additional noise removal techniques like inverse Fourier transform filtering or

by expanding the input window of the models so they observe the decay over a larger window.

Because multiple models can be deployed at once, it is also possible that a voting scheme

can be implemented between them to improve overall system performance. However, this

noise was likely generated by equipment in the lab space used for the testing and is unlikely

to happen where these devices will be deployed in the first place. Thresh 2, Thresh 3 and

LGR models had the lowest combined amount of false detections (positive and negative),

making them the leading candidates for the presence/activity index application once noise

considerations are addressed. The ANN models presented higher numbers of false detections

but they still perform considerably better than the remaining models and could also prove

useful in this application. The lower detection delay and higher capture rate of the Thresh

3 model make it the more attractive choice for this application overall.

The system also has potential applications in smart fish passage systems, where barriers

can be activated to selectively deny upstream passage in the presence of invasive species.

The sensors could also be deployed with an integrated deterrence mechanism (i.e., a stimulus

that causes pain/discomfort) in areas of high flow, where sea lamprey are likely to attach

to conserve energy during spawning migrations, as a way to make the journey to their

spawning grounds harder. These applications place a greater emphasis on fast detection

times, to engage the corresponding mechanism before the sea lamprey has a chance to move

on upstream, but there is a degree of flexibility with regard to false positives as long as

58

native species are not negatively impacted. Thresh 1, SEFR, and GNB all respond within

15 seconds of attachment, indicating they would be prime candidates for these applications,

especially with their high capture rates. The tree-based models, DT and RF, were slightly

slower but appear to be better suited for the task than the remaining models.

Several external factors, like the size of the sea lamprey and the location and length of

the attachment will affect the likelihood of detection with this system. The magnitude of the

response for this type of sensor is heavily dependent on the size of the sea lamprey attaching

to the panel as well as the location; the more the sensing area, is covered the larger the

magnitude of the response. Indeed, some of the models’ average length of missed attach-

ments are longer than the average time to detection, meaning it is likely that the response

magnitude was too small to be differentiated from the baseline’s natural variation. This may

be a considerable problem for a single-pixel sensor with limited sensing area depending on

the application, but could be solved by an extension of the detection algorithm for array

configurations, which may consider data from neighboring pixels when determining if a sea

lamprey is contacting the array. The attachments in the experiments conducted to build

the training set used in this paper lasted at least 30 seconds by design, but there is a lack

of detailed information on how long sea lampreys attach to surfaces in the wild. However,

skillful placement of the sensors may mitigate this issue somewhat; as lampreys generally

attach to surfaces to rest before continuing their journey toward spawning grounds, placing

the sensors in areas that require considerable exertion to get through could be a way to

promote the long attachments required for detection.

4.6 Summary

This chapter presents the use of simple machine learning and tunable thresholding models to

identify the characteristics of an IDE sensor’s response to sea lamprey attachment, enabling

detection of the invasive species. The average voltage and rate of change of voltage were

selected as the inputs for the detection algorithm since the characteristic exponential decay

response will also have a rate profile that is distinct from the baseline rate. A training set

59

of labeled sensor response data was constructed through experiments with live sea lamprey

on IDE sensors to train the selected models. A threshold model that could be tuned via

optimization approaches was presented as the simplest detection approach, flagging data

points within specified bounds in the input space as attachments. Artificial neural networks,

logistic regression, Gaussian Naive Bayes, decision trees, and random forest classifier models

were also explored. These were successfully deployed on a microcontroller platform interfac-

ing with an IDE sensor and achieved classification accuracy between 65-75% in live animal

validation testing. The detection performance of the models was examined and they demon-

strated high numbers of false positive and negative detections, though noise observed during

testing very likely contributed to this. The models did manage to capture the majority

of the attachments with capture rates from 63-85%. The slower, more conservative models

(Thresh 2, Thresh 3, LGR, ANNs), which report lower numbers of false detections are better

suited for monitoring applications for the sensor, whereas the faster more aggressive models

(Thresh 1, SEFR, GNB) would be a better fit for reaction/deterrence applications. Thresh 3,

which used x and y as a cost function during tuning, can be considered the best-performing

algorithm as it makes the most appearances in the top 5 performers across the performance

criteria presented here. It is possible to continue refining the presented models on a power-

ful workstation and simply follow the same deployment methodology outlined in this paper

to deploy the improved algorithms. While attachment time and location are parameters

that could affect the performance of these sensors, possible workarounds for both issues are

presented.

60

CHAPTER 5

THIRD GENERATION MEASUREMENT SYSTEM AND SENSOR DESIGN

This chapter presents improvements to the sensing system to address the sensitivity/detection

speed and sensor robustness issues presented in Chap. 4. This includes the design of the

third-generation measurement approach developed to increase the magnitude of the sensor

response through amplification to improve sensitivity and detection speed. The implementa-

tion of this system using discrete components is presented. The design of the third-generation

sensors, which aims to address robustness concerns by replacing conductive carbon paint as

electrode material with 3D-printed conductive carbon-based polymer, is also presented. The

chapter ends with a presentation of the updated fabrication method.

5.1 Third Generation Sensing System Design

A new sensor measurement approach was developed to address the shortcomings of the volt-

age divider measurement circuit, namely the inability to amplify the sensor response signal

independently from the DC bias required to observe both positive and negative changes

Figure 5.1 Circuit schematic showing the Wheatstone bridge and instrumentation amplifier-
based measurement system for a single IDE sensor. The output of the bridge is amplified
by the in-amp and offset by a DC bias generated from a voltage follower circuit.

61

in sensor resistance. Separating the signal response from the bias is required to allow for

targeted amplification, which would increase the overall sensitivity of the system while re-

taining its bidirectional measurement capabilities. A Wheatstone bridge circuit topology,

highlighted in purple in Fig. 5.1, was selected to enable this separation. This topology adds

another set of resistors, which will be referred to as the reference branch, in parallel with

the standard voltage divider that was previously used to measure the sensors, and which

will now be referred to as the sensing branch. The output of this circuit is the difference

between the node voltages in the middle of each branch, thus, the circuit not only converts

the change in impedance across the sensor to a voltage signal but also directly compares it

to a reference voltage set by the reference branch. This effectively extracts the sensor output

response, allowing it to be amplified. Using a voltage divider analysis for the circuit shown

in Fig.3.1-A, the output voltage (Vout) can be calculated as:

(cid:18)

VoutW =

Rbal2
Rbal2 + Rbal1

−

Rsense
Rsense + Rcal

(cid:19)

Vinput

(5.1)

The Wheatstone bridge is implemented using Analog Devices’ AD5242, which is a digital

potentiometer chip with two individually controllable digital potentiometers with 256 pos-

sible resistance values going from approximately zero ohms to the nominal resistance value

of the device. The resistance across the A and B terminals of each potentiometer is fixed to

the nominal value (within tolerances), and the resistance between either terminal and the

wiper is modified by sending instructions to the device. The chip is controlled using the I2C

protocol, which is compatible with Arduino devices. The first potentiometer on the chip

(Pot 0) is used to control the proportion of resistances in the reference branch of the circuit.

The second potentiometer (Pot 1) is used in rheostat configuration to adjust the reference

resistance in the sensing branch.

Texas Instruments’ INA333, a low-power, rail-to-rail instrumentation amplifier, was se-

lected as the amplification stage of the sensing system because of the topology’s high in-

put impedance, ensuring minimal disturbance to the measurement circuit. The gain is set

62

through a single external resistor and can vary from 1 to 1000, following A = 1+100kΩ/Rgain.

A 100 kΩ potentiometer was used as Rgain to allow for setting any gain within this range.

The INA333 can also apply a bias or reference voltage to the output by providing the

corresponding voltage at the Vref pin so that their output follows the following equation:

VoutA = A ∗ (V+ − V−) + Vref . This voltage must have a low source impedance to avoid af-

fecting the gain circuit, so it is provided through a voltage buffer circuit using TI’s OPA333.

The positive input of the buffer is tied to the wiper of another 100 kΩ potentiometer Rbias

with its main terminals connected between Vcc and ground, allowing for bias voltages across

the entire supply range. Additionally, for the purpose of calibrating the system, two more

voltage followers are included in the measurement system, allowing for direct measurement

of the voltage across Rsense and Rbal2.

5.2 Third-Generation Sensor Design and Fabrication

Figure 5.2 A) Schematic view of the third-generation sensor with highlighted dimensions
includes different points of view to capture 3D geometry. B) Schematic view of 3D printed
connectors for lead attachment, including relevant dimensions. C) Example of a manufac-
tured 3rd-generation IDE device.

63

Conductive carbon PLA was selected as the electrode material for the third-generation

sensing devices due to having increased robustness while maintaining similar electrical char-

acteristics. The schematic for the third-generation design is presented in Fig. 5.2-A. The

move towards 3D-printed fabrication also allowed for improvements in the design, including

a smaller gap between electrodes and a higher number of fingers per electrode. This effec-

tively increases the fill ratio of the device, which should lead to a corresponding increase in

sensitivity. The updated design features a slightly larger sensing area (40 mm x 40 mm) as

well as stubs, seen in the side view of Fig.5.2-A, that protrude from the back of the elec-

trodes to facilitate the attachment of electrical leads on the opposite side of the substrate

using a conductive 3D-printed connector presented in Fig.5.2-B. The electrode’s minimum

thickness (0.6 mm) was determined iteratively, reducing the thickness of the design until

the 3D printer failed to consistently produce a viable sensor. Trace width was selected to

optimize the fill factor of the device under the area and separation constraints. An example

of the third generation device is presented in Fig. 5.2-C.

Figure 5.3 Third-generation device manufacturing flow.

Clear acrylic panels were kept as the substrate for the updated IDE sensor to allow light

to pass through them and promote algal growth on their surfaces. Conductive carbon-based

polylactic acid (PLA) filament (Electrically Conductive Composite PLA by Protopasta) was

used as the electrode material, replacing the Marine-grade conductive carbon spray paint.

Fig. 5.3-A shows the device fabrication process beginning with the electrodes being 3D-

64

printed on a standard fused deposition modeling printer. The square acrylic panel substrate

(approximately 80 mm x 80 mm) is then drilled to allow the sensor stubs to pass through to

the opposite side for electrical lead connection. The 3D-printed electrodes are then attached

to the substrate using a waterproof adhesive. A soldering iron is then used to melt the acrylic

around the stub to ensure a watertight seal around the lead attachment area. The leads

are prepared by partially stripping an insulated wire and wrapping the exposed conductor

around the conductive 3D-printed connector, highlighted in Fig. 5.3-A, and fusing the wire

to the connector using a soldering iron. Conductive silver epoxy is then applied to the

stub before attaching the connector to ensure a robust electrical connection. Marine-grade

silicone sealant is then used to waterproof the lead attachment. Finally, the leads were

braided together to minimize the impact of ambient electrical noise on the sensor response

signal. The devices are then subjected to open-circuit and short-circuit verification tests

before proceeding to testing.

65

CHAPTER 6

CHARACTERIZATION OF SENSOR RESPONSE TO ALGAL BIOFILM
ACCUMULATION

This chapter presents the attempted characterization of the response of the sensing system

to algal biofilm accumulation to better understand how the system will perform in the field.

The methodology used for the characterization experiments using a live microalgae culture

in a controlled setting is presented. The response of the sensor voltage during the biofilm

accumulation tests and the change in sensor resistance after each test are analyzed and

compared to biomass accumulation measures to determine if there is a relationship between

either sensor output and biological accumulation on the sensor. The ability of the sensors to

detect biological contact after developing biofilms on their surface is also examined. Finally,

the implications of these results for sea lamprey detection and potential ways to address

them are discussed.

6.1 Experimental Design for Algal Biofilm Accumulation Response Character-

ization

6.1.1 Experimental Set Up

The characterization experiments took place in a UTEX 2L lab-scale photobioreactor (PBR)

paired with a UTEX RGB lighting platform. The PBR system consists of three major

components: a clear glass tank with a gas inlet connected to an aeration stone; a pump that

sends pressurized ambient air to the tank; and a lighting rig that provides full-spectrum light

for algal nourishment. To induce flow conditions that are conducive to biofilm formation,

the entire system was placed on top of a magnetic stirrer, and a 4-cm stir bar was added

to the PBR tank. A picture of the full experimental setup including the PBR, lighting rig,

and the magnetic stirrer is shown in Fig. 6.1-A. The microalgae that were cultured during

testing are Chlorella sorokiniana MSU ; it is a robust alga found in the Great Lakes region,

which is readily available. The use of native algae species will also make the experiments

more relevant to potential biofilm evolution in the field.

66

Figure 6.1 A) Experimental setup for biofilm accumulation tests, including the glass PBR
tank, RGB LED lighting rig, and magnetic stirrer, a sensor is seen through the gap in the
lighting rig. B) Sensors placed in a 3D-printed base before being introduced to the PBR
tank.

67

Figure 6.2 Circuit setup for the biofilm measurement experiment.

For each run, 3 sensors were submerged in the active culture using a circular 3D-printed

base to hold them at the bottom of the PBR tank, as seen in Fig. 6.1-B. These were

connected to the measurement system through two 16-channel multiplexers as seen in 6.2.

The multiplexers were controlled simultaneously, as each sensor had one lead connected to

each multiplexer on the same channel. This allowed each sensor to be measured individually

by addressing the proper channel on the multiplexers (Channel 1 through 3). An SD card

module was used as data storage for the system.

For the first round of testing, the polarity of the voltage across the electrodes was fixed

throughout the test (each electrode remained connected to the same node in the measure-

ment system), mirroring how the sensors were measured in previous sea lamprey detection

experiments. However, as the system uses exposed electrodes with a DC potential in an elec-

trolyte solution, there were concerns that degradation of the sensors due to electrochemical

reactions at the electrode surfaces could affect the system output. To address this, a relay

board was introduced into the system, highlighted in red in 6.2, to alternate which of the

electrodes is connected to the ground node. This effectively alternates which electrode is at

the highest electrical potential and could provide some protection from the electrochemical

degradation.

68

6.1.2 Experimental Procedure

Before each test, the PBR tank was thoroughly cleaned to remove residue from previous

cultures. The sensors were then introduced into the tank along with the stir bar before being

connected to the multiplexers. Once the sensors were connected to the measurement system,

the algal culture mixture was poured into the tank, and the magnetic stirrer was turned on.

The system then followed a two-step self-calibration procedure, first adjusting Rref to achieve

a specified baseline voltage across the IDE and then adjusting the proportion between Rbal1

and Rbal2 to match. This was done for each sensor individually, and the resulting values

of Rref and Rbal2 were recorded for later comparison. After calibrating, the system begins

measuring the sensors every 2 seconds (0.5 Hz sampling frequency). The system connects to

each device sequentially to collect a measurement, and once done, it disconnects them from

power until the next measurement time. For the second set of tests, the electrode connections

are toggled after every measurement (including during calibration). Data are recorded to

the SD card in a CSV format, and are also recorded through a serial terminal application for

redundancy. Tests were run for approximately 2-5 days to ensure varying levels of biofilm

growth on the sensors. The LED lighting rig was kept on at all times throughout each run.

After each test, the system is allowed to recalibrate to record the change in Rref values, and

an additional hour of data recording is captured to observe how the sensor baseline behaves

once the devices have been biofouled. After this, the sensors were extracted from the PBR,

and any biofilm that had accumulated on the leads or on the back of the sensing panel was

removed to ensure only the biofilm in contact with the sensor traces remained. To quantify

the amount of biofilm accumulated on the sensors during the experiment, each sensor was

weighed 4 times using a milligram scale to generate an average dry weight measurement

before the experiment and again 48 hours after the experiment was complete to allow for all

of the moisture in the biofilm to evaporate. Examples of how the sensors looked immediately

after the test and after 48 hours of drying are presented in Fig. 6.3-A and B, respectively.

After the monitoring tests were completed and the biofilm accumulation was quantified,

69

Figure 6.3 A) Representative sample of a sensor immediately after a biofilm accumulation
test. B) The same sensor after a 48-hour drying period, ready for the post-test weight
measurement.

several sensors were selected and subjected to an additional test to verify the effect of bio-

fouling on the sensor response to biological contact. Three sensors at a time were introduced

into a container with a diluted algae culture, the system was allowed to calibrate each sen-

sor, after which baseline data was recorded for 10 minutes. Once this time had passed, the

sensors were subjected to a thumb press that was sustained for 1-2 minutes, with a 1-minute

wait period before interacting with the next sensor. This interaction was repeated at least

twice during the test. After the interaction cycles were completed, the system was allowed

to record another 10 minutes of baseline data. Voltage response data were recorded in the

same fashion as the biofilm monitoring experiments.

70

Figure 6.4 Sensor voltage response plots for the 4 tests performed with the polarity of the
sensors remaining fixed: A) Fixed Polarity Test 1, B) Fixed Polarity Test 3, C) Fixed Polarity
Test 3, and D) Fixed Polarity Test 4.

6.2 Sensor Response to Biofilm Accumulation

6.2.1 Fixed Polarity Tests

The sensors’ voltage responses during the biofilm accumulation tests under fixed polarity

conditions are presented in Fig. 6.4. Note that the sensor response varies considerably

both during each test and across the set of tests. In Fig. 6.4 - A, which was the longest

test in the set, the response of the first sensor (S1) shows an initial increase in the output

voltage corresponding to a decrease in the resistance of the sensor, a subsequent decay

in the response corresponding to the opposite effect, followed finally by another period of

increasing voltage response. The response of the second sensor (S2) saturated very quickly

and decayed while presenting a high-frequency noise component of greater amplitude than

the other signals before saturating again. The third sensor (S3) response for this test also

saturated after approximately 2 days, but did not present the same level of high-frequency

71

Table 6.1 Resistance Measurements and Biomass Accumulation Results for Fixed Polarity
Tests

Test - Sensor

Initial Rref Final Rref

R∆

Biofilm Accumulation (mg)

Test 1 - Sensor 1
Test 1 - Sensor 2
Test 1 - Sensor 3
Test 2 - Sensor 1
Test 2 - Sensor 2
Test 2 - Sensor 3
Test 3 - Sensor 1
Test 3 - Sensor 2
Test 3 - Sensor 3
Test 4 - Sensor 1
Test 4 - Sensor 2
Test 4 - Sensor 3

3906.25
1953.13
6601.56
7382.81
3867.19
6562.50
5468.75
5390.63
5546.88
3046.88
3789.06
2773.44

1757.81
1562.50
2617.19
1171.89
1523.44
2187.50
3828.13
1406.25
3515.63
5781.25
4492.19
1992.19

-2148.44
-390.63
-3984.38
-6210.94
-2343.75
-4375.00
-1640.63
-3984.38
-2031.25
2734.38
703.13
-781.25

5
9
8
11
0
5
1
9
7
69
19
44

noise components when it decayed back into the ADC range at the beginning of day 3. In

Fig. 6.4 - B, which ended up being the shortest test in the set due to a power interruption,

all three sensor responses were generally increasing during the duration of the test. There

is evidence of crosstalk between S1 and S3 as their responses share sudden transients not

observed in S2 at several points in the recording. The third test, taking place over 2 days,

is presented in Fig. 6.4 - C. The sensor responses for this test present a combination of

the behaviors described for the previous tests; S1 very quickly increases to a stable baseline

before slowly decaying after approximately 1.25 days, S2 saturates after 1 day and remains at

the saturation point for the remainder of the test and S3 presents a slow increase throughout

the length of the test. The final test, presented in Fig. 6.4 - C, was stopped just short of a

3-day duration. During this test, S1 demonstrated another type of response with an initial

increase that peaks before entering what appears to be an exponential decay, while S2 and

S3 both showed an initial decrease followed by a steady increase for the remainder of the

test. The amount of biomass accumulated on each sensor for each test is presented in Table

6.1 together with the initial and final values for Rref .

The amount of biomass accumulated on the sensors during these tests was in the order

of several milligrams (average value of 15.5 mg), with only the sensors in the final test

72

consistently having more than 10 milligrams of biomass accumulated on their surface; if we

don’t consider the values for the fourth test the average drops to 6.1 mg. This discrepancy

is likely due to differences in the way the sensors were mounted for the final test. Tests 1

through 3 had the sensing surface pointed toward the center of the PBR tank, resulting in

considerable biofilm on the back of the devices and not on the sensor traces; for the final

fixed polarity test and all subsequent tests, the sensor were placed facing the perimeter of

the tank, leading to increased biofilm formation directly on the traces. Except for the final

test, the sensors generally calibrated to a lower value of Rref after the tests were completed,

indicating a decrease in the sensor resistance. The variety of responses demonstrated during

these tests, along with small amounts of biofilm accumulation achieved, suggests it will be

difficult to draw a link between biofilm accumulation and the voltage response of the sensors

for these tests.

6.2.2 Alternating Polarity Tests

The sensors’ voltage response during the biofilm accumulation tests under alternating polar-

ity conditions are presented in Fig. 6.5. It is immediately clear that the voltage responses

across this set have greater uniformity than those in the fixed polarity test set. The results

from the first test of this set, presented in Fig. 6.5-A, show the general response shapes

observed throughout the set. The response for sensor 1 (S1) shows a short increase which

peaks quickly, followed by a decay that eventually levels off and proceeds into a slow increase

in the baseline. This response pattern will be referred to as Type 1-A from here on out. S2

appears to have two distinct phases in its response, the first following the increasing form of

an exponential decay function and settling to a stable baseline for around 1.5 days, followed

by a decay period. This response pattern will be referred to as Type 1-B for the remainder

of this work. In contrast, S3 follows the increasing form exponential decay for the duration

of the test, which will hereby be referred to as Type 2. The next test, the results of which

are presented in Fig. 6.5-B, mirrors these response shapes with S1 and S3 following the Type

1-B pattern and S2 following the Type 2 pattern. Similarly, the results for test 3, which are

73

Figure 6.5 Sensor voltage response plots for the 4 tests performed while switching the polarity
of the sensors: A) Alternating Polarity Test 1, B) Alternating Polarity Test 3, C) Alternating
Polarity Test 3 and D) Alternating Polarity Test 4.

presented in Fig. 6.5-C, continue this pattern with S1 following the Type 1-A response, S2

following the Type 2 response, and S3 following the Type 1-B response. Finally, the results

of the last test in the alternating polarity set, presented in Fig. 6.5-D, replicate the pattern

seen in test 3 with the exception of S3 having a less pronounced decay. The amount of

biomass accumulated on each sensor for each test is presented in Table 6.2 together with the

initial and final values for Rref .

Greater levels of biofilm accumulation were achieved for the alternating polarity test

set than for the fixed polarity set due to the change in sensor mounting position, with an

average biomass accumulation of 24.9 mg versus an average of 15.5 mg for the fixed test

set. Additionally, unlike the fixed polarity test set, there is an apparent pattern between the

voltage response types and whether the sensors calibrate to higher or lower values of Rref .

The grouping of the sensor voltage responses by whether they correspond to an increase or

74

Table 6.2 Resistance Measurements and Biomass Accumulation Results for Alternating Po-
larity Tests

Test - Sensor

Initial Rref Final Rref

R∆

Biofilm Accumulation (mg)

Test 1 - Sensor 1
Test 1 - Sensor 2
Test 1 - Sensor 3
Test 2 - Sensor 1
Test 2 - Sensor 2
Test 2 - Sensor 3
Test 3 - Sensor 1
Test 3 - Sensor 2
Test 3 - Sensor 3
Test 4 - Sensor 1
Test 4 - Sensor 2
Test 4 - Sensor 3

17968.75
29296.875
28906.25
42968.75
50000
34375.0
36718.75
44921.875
35546.875
19531.25
33203.125
35937.5

39062.5
31640.625
14062.5
43359.375
23046.875
46484.375
53906.25
10937.5
59375.0
51953.125
15234.375
29296.875

21093.75
2343.75
-14843.75
390.625
-26953.125
12109.375
17187.5
-33984.375
23828.125
32421.875
-17968.75
-6640.625

29
20
46
6
35
9
24
29
49
22
38
13

decrease Rref and is presented in Fig. 6.6-A and B, respectively.

The response types most commonly associated with increases in resistance are those

described by Type 1-A and Type 1-B responses as seen in 6.6-A, while the Type 2 response

shape was generally associated with a decrease in Rref . The only observed exception to this

trend is the S3 response from test 4, which presents a Type 1-B response and the smallest

negative change in Rref for the set. This hints at a complex relationship between biomass

accumulation, the voltage response, and the change in resistance across the sensors.

6.3 Resistance Change versus Accumulated Biomass Study

Given the variety of voltage responses observed for similar amounts of biofouling, it may be

useful to study the relationship between the net resistance change occurring in the sensors

and the amount of biomass accumulated on them. For the fixed polarity test set, biofilm

accumulation as a function of the change in resistance is shown in Fig. 6.7 - A. While

there appears to be a weak positive correlation between the change in resistance and biofilm

accumulation for this set, with a correlation factor of 0.489, this result is skewed due to the

results from the final test having considerably more biomass than those from the first 3 tests.

If these values are excluded, the correlation factor drops to 0.210, and the resulting linear

75

Figure 6.6 A) Combination plot including all the sensor voltage responses for the alternating
polarity test set corresponding to an increase in Rref . B) Combination plot including all the
sensor voltage responses for the alternating polarity test set corresponding to a decrease in
Rref .

relationship approximates a flat line, indicating little to no relationship between the test

variables. Similarly, for the results of the study using the data from the alternating polarity

test, shown in Fig. 6.7-B, the linear correlation coefficient is low (R2 = 0.151), indicating a

poor relationship between biomass accumulation and change in resistance for this test set.

The data for this test set does appear to be grouped into two clusters corresponding to the

76

Figure 6.7 A) Change in Rref plotted against the total amount of biomass accumulated on
the sensors for the fixed polarity test set. B) Change in Rref plotted against the total amount
of biomass accumulated on the sensors for the alternating polarity test set.

response shapes with the Type 1-A and 1-B responses from approximately -700 to 3000 ohms

and the Type 2 responses from -3500 to -1500 ohms. However, the small sample size makes

it difficult to generalize a relationship from these two clusters.

6.4 Post-fouling Biological Contact

The sensors presenting the highest changes in resistance for the alternating polarity test set

(increased resistance: Test 3 S1, Test 3 S3, Test 4 S1, decreased resistance: Test 2 S2, Test 3

S2, Test 4 S2) were selected for an additional test to observe how the biofouling would affect

their response biological contact. The results from these experiments for the sensors with

increased resistance are presented in Fig. 6.8 - A and B. In these plots S1 corresponds to

Test 3’s S1 (R∆ = 44RDAC units), S2 to Test 4’s S1 (R∆ = 83RDAC units) and S3 to Test 3

S3 (R∆ = 61RDAC units). In the case of S1, for both tests, the biological contact is registered

in the response as a sudden peak followed by decay with the baseline trending downwards

as the test progressed. During both tests, the response from S2 is markedly noisier than

the other two sensors, making it harder to distinguish the contact response. This suggests

more aggressive filtering may be required as the change in resistance increases. The contact

response is also inconsistent for this sensor, presenting as an increase in the response in

Contact Test 1 and a decrease in Contact Test 2. On the other hand, S3 presents a sharp

increase in the sensor output, which holds through the entirety of the contact. There is a

77

Figure 6.8 A) Sensor responses in the initial biological contact test using the sensors with
the highest increased resistance. B) Sensor responses in follow-up biological contact test
using the sensors with the highest increased resistance. C) Sensor responses in the initial
biological contact test using the sensors with the highest decrease in resistance. D) Sensor
responses in follow-up biological contact test using the sensors with the highest decrease in
resistance.

sudden transient in S3 towards the end of the first test, which can be explained by established

biofilm breaking off as the touch is released.

The results from the experiments using the sensors with decreased resistance are presented

in Fig. 6.8 - C and D. In these plots S1 corresponds to Test 2’s S2 (R∆ = −69RDAC units),

S2 to Test 4’s S2 (R∆ = −46RDAC units) and S3 to Test 3 S2 (R∆ = −89RDAC units).

The response of the first sensor in these tests, S1, presents a similar response to biological

contact as the first sensor in Contact Test 1 and 2, but with a smaller response magnitude.

ContactS2 shows the clearest response to biological contact out of the three sensors with

decreased resistance, but its response also suffered from inconsistency between Contact Test

3 and 4. Similar to ContactS1, the response of ContactS3 to biological contact presents a

78

small magnitude response to biological contact in Contact Test 3, but it becomes almost

imperceptible in Contact Test 4.

6.5 Discussion

The targeted deployment environment of the automated sea lamprey sensing system in the

tributaries of the Great Lakes will be conducive to the accumulation of algal and bacterial

biofilm on the devices. As was shown in this study, environmental conditions that favor

biofilm growth will have a clear effect on the baseline of the sensors that will need to be

addressed before field deployment is viable. This effect has proven difficult to characterize

due to the wide range of responses observed throughout both test sets performed in this work.

One potential explanation is the presence of a non-biological contribution to the change in

resistance across the sensor. The algal growth medium is much more conductive than the

water the sensors are typically tested in ( 4000 S/cm vs 180-230 S/cm), which makes

it easier for electrochemical reactions to occur on the electrode surfaces. These reactions

have the potential to alter the electrode chemistry, depleting the electrode that is held at a

positive potential (anode) of its conductive carbon through oxidation. Indeed, for some of

the sensors in the fixed polarity test set had one of their electrodes rendered non-conductive,

effectively disabling the device. At the same time, it is important to note that this is at odds

with the pattern observed for the change in calibrated resistance after the fixed polarity

tests, which generally presented a reduction in Rref that should correspond to a decrease

in the sensor resistance as well. This decrease should not be possible if one of the sensing

electrodes was completely depleted. However, the sensors are slightly porous, which may

allow for the highly conductive medium to penetrate the sensors, establishing an electrical

connection with the electrical lead that results in a lower resistance. A potential indicator

that alternating the polarity of the sensors is protective against electrochemical depletion is

given by the fact that while some sensors in the alternating test set also suffered degradation,

it was not as extreme as with the sensor tests under fixed polarity.

Even with a reduction in the effects of electrochemical reactions on the device resistance

79

in the alternating polarity tests, the sensor responses for these tests did not present a con-

sistent pattern with regard to algal biofilm accumulation. The composition of the biofilm

may contribute to this observed behavior as they have two major components, microalgal

cells and extracellular polymeric substances (EPS), which contribute to the physical and

electrical properties of the biofilm. While microalgae cells are insulators and should there-

fore contribute to an increase in resistance as they accumulate on the device, EPS has a

wide range of components: proteins, phospholipids, polysaccharides, nucleic acids, humic

substances, uronic acids, and some functional groups (e.g. phosphoric, carboxylic, hydroxyl

and amino groups) and they can have varying contributions to the impedance of the biofilm.

Additionally, where the biofilm accumulates on the sensors may also play a role in the

impedance contribution. Deposition directly on the sensor traces is expected to increase the

resistance observed by the measurement system, as the insulating components of the biofilm

will restrict current flow to the medium, while deposition in the gap between electrodes would

likely decrease the resistance if the EPS provides a high conductivity channel between the

two electrodes. This indicates that to accurately characterize the sensor response to biofilm

attachment, some type of visual tracking method would have to be implemented alongside

the dry weight measurement approach.

6.5.1

Implications for Field Deployment in Sea Lamprey Detection Application

Looking at the response of the fouled sensors to biological contact, it is clear that biofilm

buildup and the changes in sensor resistance observed after the tests affect their response.

While the sensors still present a reaction to biological contact when the resistance increases

after the tests, it is less consistent than what is expected of the devices, which will require

changes in the detection algorithms used alongside this system. In the case of the sensors that

presented decreased resistance after the tests, the magnitude of the response was generally

smaller, making it harder to distinguish from baseline. This could be addressed by increasing

the amplification factor of the measurement system, which was at its lowest value for all of

the tests presented in this chapter, in response to changes in Rref . Another way to address

80

these impacts is to have the next generation of detection algorithms include responses from

deteriorated sensors in their training data, and include the value of Rref as an additional

input to the algorithms.

Moving forward, all sea lamprey sensing systems based on the IDE design should adopt

the alternating polarity approach to reduce the impact of electrochemical reactions on the

devices. Another thing to note is that these experiments were run over a longer timeframe

than previous sea lamprey detection experiments, but the system was only calibrated at

the beginning and end of each test. Electrochemical reactions are driven by a potential

difference, and since the voltage across the sensors is a function of Rref it is possible that by

frequently calibrating the system, the effects of these reactions could be further reduced. The

rapid change in voltage at the beginning of most tests would indicate that the system should

be recalibrated within hours of the previous calibration. Additionally, given the impact that

the change in resistance has on the sensor’s ability to detect biological contact, Rref should

be periodically logged and reported to serve as a measure for determining whether the sensor

should be retired. It is also possible to reduce the potential difference between electrodes by

altering the bridge circuit used to measure the devices by introducing additional resistors in

series before and after the bridge.

6.6 Future Work

The complexity of the impedance contribution of the biofilm, along with its geometric de-

pendency, will require a direct impedance measurement with an AC probing signal alongside

a visual tracking component to be properly characterized as a function of sensing area cov-

erage. Single frequency impedance analysis has been used successfully before to characterize

biofilms [61][62], and suggests that relatively low-frequency signals could be used for this

purpose. This measurement can be performed without the use of expensive equipment by

using a direct digital synthesis module to generate a sine wave at a particular frequency and

using a current sensing amplifier to measure the current across a sensing resistor, but would

require a study to be performed to find the optimal probing frequency.

81

In place of a comprehensive characterization of the system’s response to biofilm accumu-

lation, another potential direction of future research is to find ways to reduce the influence

of biofilm growth conditions on the sensor baseline. This could be done using the same setup

as the biofilm accumulation and following the suggestions laid out in the previous section.

For this potential study, and future field deployments, it would be useful to include addi-

tional environmental sensors alongside the system to track variables that could also affect

the sensor response, such as temperature, pH, and water conductivity. Finally, the response

of the sensor to ”wild” biofilm accumulation in field deployment conditions should also be

studied, as it is possible that the highly conductive environment of the PBR could be a

greater contributor to the observed responses than the biofilm accumulation itself.

82

CHAPTER 7

SUMMARY

7.1 Summary of Contributions

This work presents the development of an automated sea lamprey detection system that could

become a powerful tool for monitoring these invasive fish. The system was built around low-

cost conductive carbon electrodes to facilitate deployment at scale. The methodologies used

to design, fabricate, and validate the sensors are one of the key takeaways from this work,

which could be applied to similar underwater sensing systems. Similarly, the development

of measurement and signal conditioning circuits used as a part of this system also have

potential applications in the measurement of other resistive sensors. The integration of

machine learning models with the system, which is operated by a simple microcontroller

unit, to enable automatic real-time sea lamprey detection demonstrates a novel application

of embedded machine learning. This effort also resulted in a guide for the deployment of

simple machine learning models to embedded systems. This work took the first steps towards

characterizing the sensor response to biofilm accumulation, which has the potential to enable

monitoring applications in a range of fields.

7.1.1 Problems Solved in this Thesis

This work addresses the following:

• The interdigitated electrode structure was shown to be viable for underwater contact

sensing using relative conductivity differences as a driver for contact source character-

ization.

• The response of the sensor to sea lamprey attachment using a simple voltage divider

measurement circuit was characterized

• The correlation between characterized response parameters and sea lamprey size pa-

rameters was explored

83

• Machine learning models were trained on sea lamprey response data and deployed to

a resource-constrained device to enable automatic detection of sea lamprey

• Shortcomings in the initial measurement system were addressed through an updated

measurement approach

• The effect of biofilm accumulation on the sensors was studied, and strategies to deal

with this impact were proposed

84

BIBLIOGRAPHY

[1] F. Guy, “Sea lamprey wounds (Petromyzon marinus) on a Chinook salmon (On-
corhynchus tshawytscha), St. Mary’s River-North Channel, Garden River First Nation.”

[2] “CC BY-SA 3.0 Deed | Attribution-ShareAlike 3.0 Unported | Creative Commons.”

[3] C. B. Renaud and P. A. Cochran, “Post-metamorphic Feeding in Lampreys,” in Lam-
preys: Biology, Conservation and Control: Volume 2 (M. F. Docker, ed.), Fish & Fish-
eries Series, pp. 247–285, Dordrecht: Springer Netherlands, 2019.

[4] R. L. Eshenroder, “The Role of the Champlain Canal and Erie Canal as Putative Cor-
ridors for Colonization of Lake Champlain and Lake Ontario by Sea Lampreys,” Trans-
actions of the American Fisheries Society, vol. 143, pp. 634–649, May 2014. Publisher:
Taylor & Francis eprint: https://doi.org/10.1080/00028487.2013.879818.

[5] V. C. Applegate, “Natural history of the sea lamprey, Petromyzon marinus, in Michi-

gan,” Report 55, 1950.

[6] B. R. Smith and J. J. Tibbles, “Sea Lamprey ( Petromyzon marinus ) in Lakes Huron,
Michigan, and Superior: History of Invasion and Control, 1936–78,” Canadian Journal
of Fisheries and Aquatic Sciences, vol. 37, pp. 1780–1801, Nov. 1980.

[7] W. J. Christie, “Changes in the Fish Species Composition of the Great Lakes,” Journal
of the Fisheries Board of Canada, May 1974. Publisher: NRC Research Press Ottawa,
Canada.

[8] A. Muir, C. Krueger, and M. Hansen, “Re-establishing lake trout in the Laurentian
Great Lakes: past, present, and future,” in Great Lakes fisheries policy and management:
a binational perspective, pp. 533–588, Jan. 2013. Journal Abbreviation: Great Lakes
fisheries policy and management: a binational perspective.

[9] P. J. Hrodey, S. A. Lewandoski, W. P. Sullivan, J. M. Barber, K. A. Mann, B. Paudel,
and M. J. Symbal, “Evolution of the sea lamprey control barrier program: The impor-
tance of lowermost barriers,” Journal of Great Lakes Research, vol. 47, pp. S285–S296,
Dec. 2021.

[10] N. S. Johnson, H. T. Thompson, C. Holbrook, and J. A. Tix, “Blocking and guid-
ing adult sea lamprey with pulsed direct current from vertical electrodes,” Fisheries
Research, vol. 150, pp. 38–48, Feb. 2014.

85

[11] N. S. Johnson, B. Snow, T. Bruning, and A. Jubar, “A seasonal electric barrier blocks
invasive adult sea lamprey (Petromyzon marinus) and reduces production of larvae,”
Journal of Great Lakes Research, vol. 47, pp. S310–S319, Dec. 2021.

[12] W. P. Sullivan, D. P. Burkett, M. A. Boogaard, L. A. Criger, C. E. Freiburger, T. D.
Hubert, K. G. Leistner, B. J. Morrison, S. M. Nowicki, S. N. P. Robertson, A. K.
Rowlinson, B. J. Scotland, and T. B. Sullivan, “Advances in the use of lampricides to
control sea lampreys in the Laurentian Great Lakes, 2000–2019,” Journal of Great Lakes
Research, vol. 47, pp. S216–S237, Dec. 2021.

[13] R. A. Bergstedt and M. B. Twohey, “Research to Support Sterile-male-release and Ge-
netic Alteration Techniques for Sea Lamprey Control,” Journal of Great Lakes Research,
vol. 33, pp. 48–69, Jan. 2007.

[14] “Great Lakes Fishery Commission - The Fishery.”

[15] J. V. Adams, J. M. Barber, G. A. Bravener, and S. A. Lewandoski, “Quantifying Great
Lakes sea lamprey populations using an index of adults,” Journal of Great Lakes Re-
search, vol. 47, pp. S335–S346, Dec. 2021.

[16] S. A. Lewandoski and T. O. Brenden, “Forecasting suppression of invasive sea lamprey
in Lake Superior,” Journal of Applied Ecology, vol. 59, no. 8, pp. 2023–2035, 2022.
eprint: https://onlinelibrary.wiley.com/doi/pdf/10.1111/1365-2664.14203.

[17] N. Schloesser, “Correlating sea lamprey density with environmental DNA detections in

the lab,” Management of Biological Invasions, vol. 9, no. 4, pp. 483–495, 2018.

[18] T. D. Gingera, T. B. Steeves, D. A. Boguski, S. Whyard, W. Li, and M. F. Docker,
“Detection and identification of lampreys in Great Lakes streams using environmental
DNA,” Journal of Great Lakes Research, vol. 42, pp. 649–659, June 2016.

[19] E. L. McCann, N. S. Johnson, P. J. Hrodey, and K. L. Pangle, “Characterization
of Sea Lamprey Stream Entry Using Dual-Frequency Identification Sonar,” Transac-
eprint:
tions of the American Fisheries Society, vol. 147, no. 3, pp. 514–524, 2018.
https://onlinelibrary.wiley.com/doi/pdf/10.1002/tafs.10052.

[20] G. D. Bixler and B. Bhushan, “Biofouling: lessons from nature,” Philosophical Transac-
tions of the Royal Society A: Mathematical, Physical and Engineering Sciences, vol. 370,
pp. 2381–2417, May 2012. Publisher: Royal Society.

[21] J. L. Harding and M. M. Reynolds, “Combating medical device fouling,” Trends in

Biotechnology, vol. 32, pp. 140–146, Mar. 2014.

86

[22] F. Di Pippo, L. Di Gregorio, R. Congestri, V. Tandoi, and S. Rossetti, “Biofilm growth
and control in cooling water industrial systems,” FEMS Microbiology Ecology, vol. 94,
May 2018.

[23] O. Zeriouh,

J. V. Reinoso-Moreno, L. L´opez-Rosales, M. d. C. Cer´on-
Garc´ıa, A. S´anchez-Mir´on, F. Garc´ıa-Camacho, and E. Molina-Grima, “Biofoul-
ing in photobioreactors for marine microalgae,” Critical Reviews in Biotechnol-
eprint:
ogy, vol. 37, pp. 1006–1023, Nov. 2017.
https://doi.org/10.1080/07388551.2017.1299681.

Publisher: Taylor & Francis

[24] A. M. Curtin and H. L. Buckley,

“Biofouling detection methods

widely applicable and useful across disciplines:
vol. 37, pp. 494–505, May 2021.
Publisher:
https://doi.org/10.1080/08927014.2021.1926998.

that are
a mini-review,” Biofouling,
eprint:
Taylor & Francis

[25] Q. Xie, J. Pan, C. Ma, and G. Zhang, “Dynamic surface antifouling: mechanism and
systems,” Soft matter, vol. 15, pp. 1087–1107, Feb. 2019. Place: England Publisher:
Royal Society of Chemistry.

[26] I. Gonz´alez-Afanador, H. Shi, C. Holbrook, X. Tan, and N. Sep´ulveda, “Invasive Sea
Lamprey Detection and Characterization Using Interdigitated Electrode (IDE) Contact
Sensor,” IEEE Sensors Journal, vol. 21, pp. 27947–27956, Dec. 2021. Conference Name:
IEEE Sensors Journal.

[27] I. Gonz´alez-Afanador, C. Chen, G. Morales-Torres, S. Meihls, H. Shi, X. Tan, and
N. Sep´ulveda, “Real-time invasive sea lamprey detection using machine learning clas-
sifier models on embedded systems,” Neural Computing and Applications, vol. 36,
pp. 16195–16212, Sept. 2024.

[28] H. Shi, M. Al-Rubaiai, C. M. Holbrook, J. Miao, T. Pinto, C. Wang, and X. Tan,
“Screen-Printed Soft Capacitive Sensors for Spatial Mapping of Both Positive and Neg-
ative Pressures,” Advanced Functional Materials, vol. 29, no. 23, p. 1809116, 2019.
eprint: https://onlinelibrary.wiley.com/doi/pdf/10.1002/adfm.201809116.

[29] H. Shi, I. Gonz´alez-Afanador, C. Holbrook, N. Sep´ulveda, and X. Tan, “Soft Pres-
sure Sensor for Underwater Sea Lamprey Detection,” IEEE Sensors Journal, vol. 22,
pp. 9932–9944, May 2022. Conference Name: IEEE Sensors Journal.

[30] Y. Cao, H. Shi, X. Tan, and N. Sep´ulveda, “Enabling Negative Pressure Sensing
Through Ferroelectret Device,” IEEE Sensors Letters, vol. 6, pp. 1–4, Aug. 2022. Con-
ference Name: IEEE Sensors Letters.

87

[31] Y. Duan, S. He, J. Wu, B. Su, and Y. Wang, “Recent Progress in Flexible Pressure

Sensor Arrays,” Nanomaterials, vol. 12, p. 2495, July 2022.

[32] H. Shi, C. M. Holbrook, Y. Cao, N. Sep´ulveda, and X. Tan, “Measurement of suc-
tion pressure dynamics of sea lampreys, Petromyzon marinus,” PLOS ONE, vol. 16,
p. e0247884, Apr. 2021. Publisher: Public Library of Science.

[33] X. Tan, “E-Skin Technology Could Aid in Sea Lamprey Fight,” Jan. 2020.

[34] C. Chen, H. Shi, I. Gonz´alez-Afanador, N. Sep´ulveda, and X. Tan, “Rapid Fabrication
of Flexible Pressure Sensor Array,” IEEE Sensors Letters, vol. 7, pp. 1–4, June 2023.
Conference Name: IEEE Sensors Letters.

[35] H. Shi, Y. Mei, I. Gonz´alez-Afanador, C. Chen, S. Miehls, C. Holbrook, N. Sep´ulveda,
and X. Tan, “Automated Soft Pressure Sensor Array-Based Sea Lamprey Detection
Using Machine Learning,” IEEE Sensors Journal, vol. 23, pp. 7546–7557, Apr. 2023.
Conference Name: IEEE Sensors Journal.

[36] J.-G. Guan, Y.-Q. Miao, and Q.-J. Zhang, “Impedimetric biosensors,” Journal of Bio-

science and Bioengineering, vol. 97, no. 4, pp. 219–226, 2004.

[37] N. Hikmah, H. F. Hawari, and M. Gupta, “Design and simulation of interdigitated
electrode for Graphene-SnO2 sensor on acetone gas,” Indonesian Journal of Electrical
Engineering and Computer Science, vol. 19, pp. 119–125, July 2020. Number: 1.

[38] L. Wang, M. Veselinovic, L. Yang, B. J. Geiss, D. S. Dandy, and T. Chen, “A sensitive
DNA capacitive biosensor using interdigitated electrodes,” Biosensors & Bioelectronics,
vol. 87, pp. 646–653, Jan. 2017.

[39] J. Wang, C. Wu, N. Hu, J. Zhou, L. Du, and P. Wang, “Microfabricated Electrochemical
Cell-Based Biosensors for Analysis of Living Cells In Vitro,” Biosensors, vol. 2, pp. 127–
170, June 2012. Number: 2 Publisher: Molecular Diversity Preservation International.

[40] P. Van Gerwen, W. Laureyn, W. Laureys, G. Huyberechts, M. Op De Beeck, K. Baert,
J. Suls, W. Sansen, P. Jacobs, L. Hermans, and R. Mertens, “Nanoscaled interdigitated
electrode arrays for biochemical sensors,” Sensors and Actuators B: Chemical, vol. 49,
pp. 73–80, June 1998.

[41] J. S. Daniels and N. Pourmand, “Label-Free Impedance Biosensors: Opportuni-
eprint:

ties and Challenges,” Electroanalysis, vol. 19, no. 12, pp. 1239–1257, 2007.
https://analyticalsciencejournals.onlinelibrary.wiley.com/doi/pdf/10.1002/elan.200603855.

88

[42] C. Gaul, “Lampetra appendix (American brook lamprey).”

[43] B. Karvel-Fuller, “Ichthyomyzon fossor (Northern brook lamprey).”

[44] T. Acciaioli, “Ichthyomyzon castaneus (Chestnut lamprey).”

[45] C. Blumbergs, “Ichthyomyzon unicuspis (Silver lamprey).”

[46] A. Dizon and M. E. Orazem, “On the impedance response of interdigitated electrodes,”

Electrochimica Acta, vol. 327, p. 135000, Dec. 2019.

[47] S. Cherry, “Petromyzon marinus (Eel sucker).”

[48] N. Ha, K. Xu, G. Ren, A. Mitchell, and J. Z. Ou, “Machine Learning-Enabled Smart
Sensor Systems,” Advanced Intelligent Systems, vol. 2, no. 9, p. 2000063, 2020. eprint:
https://onlinelibrary.wiley.com/doi/pdf/10.1002/aisy.202000063.

[49] H. Keshavarz, M. S. Abadeh, and R. Rawassizadeh, “SEFR: A Fast Linear-Time Clas-

sifier for Ultra-Low Power Devices,” Nov. 2020. arXiv:2006.04620 [cs, stat].

[50] V. Hurbungs, V. Bassoo, and T. P. Fowdur, “An enhanced binary classifier for Edge

devices,” Microprocessors and Microsystems, vol. 93, p. 104596, Sept. 2022.

[51] F. Shabani, H. Philamore, and F. Matsuno, “An Energy-Autonomous Chemical Oxygen
Demand Sensor Using a Microbial Fuel Cell and Embedded Machine Learning,” IEEE
Access, vol. 9, pp. 108689–108701, 2021. Conference Name: IEEE Access.

[52] B. R. Ibrahim, F. M. Khalifa, S. R. M. Zeebaree, N. A. Othman, A. Alkhayyat, R. R. Ze-
bari, and M. A. M. Sadeeq, “Embedded System for Eye Blink Detection Using Machine
Learning Technique,” in 2021 1st Babylon International Conference on Information
Technology and Science (BICITS), pp. 58–62, Apr. 2021.

[53] B. da Silva, A. W. Happi, A. Braeken, and A. Touhafi, “Evaluation of Classical Ma-
chine Learning Techniques towards Urban Sound Recognition on Embedded Systems,”
Applied Sciences, vol. 9, p. 3885, Jan. 2019. Number: 18 Publisher: Multidisciplinary
Digital Publishing Institute.

[54] C. Huang and L. Jiang, “Data monitoring and sports injury prediction model based on
embedded system and machine learning algorithm,” Microprocessors and Microsystems,
vol. 81, p. 103654, Mar. 2021.

89

[55] F. Pedregosa, G. Varoquaux, A. Gramfort, V. Michel, B. Thirion, O. Grisel, M. Blondel,
P. Prettenhofer, R. Weiss, V. Dubourg, J. Vanderplas, A. Passos, D. Cournapeau,
M. Brucher, M. Perrot, and Duchesnay, “Scikit-learn: Machine Learning in Python,”
Journal of Machine Learning Research, vol. 12, no. 85, pp. 2825–2830, 2011.

[56] G. Louw, “glouw/tinn,” Apr. 2023. original-date: 2018-03-27T07:16:59Z.

[57] S. Salerno, “MicroMLGen for Python,” Oct. 2020.

[58] K. Deb, A. Pratap, S. Agarwal, and T. Meyarivan, “A fast and elitist multiobjective
genetic algorithm: NSGA-II,” IEEE Transactions on Evolutionary Computation, vol. 6,
pp. 182–197, Apr. 2002. Conference Name: IEEE Transactions on Evolutionary Com-
putation.

[59] A. P. Wierzbicki, “The Use of Reference Objectives in Multiobjective Optimization -
Theoretical Implications and Practical Experience,” Aug. 1979. Num Pages: 32 Place:
IIASA, Laxenburg, Austria Publisher: WP-79-066.

[60] J. Blank and K. Deb, “Pymoo: Multi-Objective Optimization in Python,” IEEE Access,

vol. 8, pp. 89497–89509, 2020. Conference Name: IEEE Access.

[61] M. McGlennen, M. Dieser, C. M. Foreman, and S. Warnat, “Using electrochemical
impedance spectroscopy to study biofilm growth in a 3D-printed flow cell system,”
Biosensors and Bioelectronics: X, vol. 14, p. 100326, Sept. 2023.

[62] M. McGlennen, M. Dieser, C. M. Foreman, and S. Warnat, “Monitoring biofilm growth
and dispersal in real-time with impedance biosensors,” Journal of Industrial Microbiol-
ogy & Biotechnology, vol. 50, p. kuad022, Aug. 2023.

90

APPENDIX

Remaining Experimental Results
This section contains complementary sensor response plots for experiments, which were not
presented in the main manuscript. Experiments are named using the lamprey’s identifier tag
and the test number for that lamprey and are presented in the order they were performed.
Plots are accompanied by a brief commentary on the experiment.
In all the plots, sea
lamprey attachments are identified with a pair of symbols: green triangles indicate when the
sea lamprey attaches to the sensor and red dots indicate when it releases; each triangle-circle
pair denotes one attachment.

Figure A-1 Sensor response for Blue-002 Test 1.

Three short attachments were observed before a sustained attachment. Noticeable expo-

nential decay observed; response was included in the correlation study.

Figure A-2 Sensor response for White-041 Test 1.

Single sustained attachment observed. A transient spike followed by noticeable exponen-

tial decay observed; response was included in the correlation study.

91

Figure A-3 Sensor response for Orange-025 Test 1.

Three short attachments were observed before a sustained attachment. Transient spike
followed by noticeable exponential decay observed; response was included in the correlation
study.

Figure A-4 Sensor response for White-037 Test 1.

Single sustained attachment observed. Transient spike followed by noticeable exponential

decay observed; response was included in the correlation study.

92

Figure A-5 Sensor response for Blue-024 Test 1.

One attachment with noticeable exponential decay observed; included in the correlation
study. Transients from 150- 250 s, possibly due to movement of electric leads, not considered
for exponential fit.

Figure A-6 Sensor response for Blue-009 Test 1.

Two short attachments observed before a sustained attachment. Noticeable exponential

decay observed; response was included in the correlation study.

93

Figure A-7 Sensor response for Pink-035 Test 1.

Single sustained attachment observed. Small transient spike followed by noticeable ex-

ponential decay observed; response was included in the correlation study.

Figure A-8 Sensor response for Orange-035 Test 1.

Single sustained attachment observed. Transient spike followed by noticeable exponential

decay observed; response was included in the correlation study.

94

Figure A-9 . Sensor response for Orange-017- Test 1.

Three short attachments observed. Last attachment lasted 10 seconds, observed transient
corresponds to lamprey rotating its oral disc on the sensor. Response was not included in
the correlation study.

Figure A-10 Sensor response for Orange-004 Test 1.

Six short attachments observed during test. Penultimate attachment lasted 13 seconds,
shows apparent decay but falls within noise variation. This response was not included in the
correlation study.

95

Figure A-11 Sensor response for Blue-018 Test 1.

Six short attachments observed during test. This response was not included in the cor-

relation study.

Figure A-12 Sensor response for White-028 Test 1.

Four short attachments observed during test. During last attachment, lasing 19 seconds,
lamprey can be observed sliding on sensor. Response shape is not exponential, it was not
included in the correlation study.

96

Figure A-13 Sensor response for Pink-034 Test 1.

Six short attachments observed during test. Last attachment lasted 11 seconds, shows
apparent decay but falls within noise variation. This response was not included in the
correlation study.

Figure A-14 Sensor response for Pink-029 Test 1.

Single sustained attachment observed. Small transient spike followed by noticeable ex-
ponential decay observed. Lamprey was only partially attached; response was not included
in the correlation study.

97

Figure A-15 Sensor response for Orange-002 Test 1.

Four short attachments observed before a sustained attachment. Noticeable exponential
decay observed, but error is an order of magnitude greater than every other fit, not included
in the correlation study.

Figure A-16 Sensor response for White-043 Test 1.

Two short attachments observed before a sustained attachment. Noticeable exponential

decay observed; response was included in the correlation study.

98

Figure A-17 Sensor response for Orange-026 Test 1.

Single sustained attachment observed. Small transient spike followed by noticeable ex-
ponential decay observed. Lamprey was only partially attached; response was not included
in the correlation study.

Figure A-18 Sensor response for Orange-017 Test 2.

Ten short attachments observed. First four attachments show apparent decay, but fre-
quency of attachment makes them difficult to distinguish from noise. Response not used in
correlation study.

99

Figure A-19 Sensor response for Orange-035 Test 2.

Five short attachments observed. Most show apparent decay but fall within noise varia-

tion. Response was not utilized in correlation study.

Figure A-20 Sensor response for White-041 Test 2.

Twelve short attachments observed. Most show apparent decay, but frequency of attach-
ment makes them difficult to distinguish from noise. Response was not utilized in correlation
study.

100

Figure A-21 Sensor response for Blue-002 Test 2.

7 short attachments observed. 3rd and 7th attachment ( 11 and 8 seconds respectively)
show apparent decay but fall within noise variation. Response was not utilized in correlation
study.

Figure A-22 Sensor response for Orange-025 Test 2.

Two short attachments showing apparent decay before sustained attachment observed.
First attachment ( 12 seconds) had a low correlation with the exponential fit, only the 3rd
was included in the correlation study.

101

Figure A-23 Sensor response for Pink-034 Test 2.

Three short attachments before sustained attachment. Transients between 100-220s,
possibly due to movement of electric leads, not considered for curve fit. Excluded from
correlation study for this reason.

Figure A-24 Sensor response for White-037 Test 2.

Single sustained attachment observed. Noticeable exponential decay observed; response

was included in the correlation study.

102

Figure A-25 Sensor response for Pink-028 Test 2.

One short attachment before a sustained attachment observed. Noticeable exponential

decay observed; response was included in the correlation study.

Figure A-26 Sensor response for Blue-024 Test 2.

One short attachment before sustained attachment observed. Noticeable exponential

decay observed; response was included in the correlation study.

103

Figure A-27 Sensor response for Pink-027 Test 2.

Ten short attachments observed. 3rd,4th, 6th, 9th and 10th attachments show apparent
decay. The largest decay appears almost linear, exponential fit correlation was low. Response
not used in correlation study.

104

Representative Response Videos
This section contains videos corresponding to the three representative tests presented in the
main manuscript. The videos show the test proceedings and include the real-time sensor
response plot in the upper left corner, as well as text annotations in the lower left corner
indicating attachment status.

Video 1. Sensor Response to Sustained Attachment, Pink-027 Test 1 - Fig. 3.3
Video 2. Sensor Response to Short Attachment, Pink-028 Test 1 - Fig. 3.4
Video 3. Sensor Response to Sliding Attachment, Orange-040 Test 1 - Fig. 3.5

Example Validation Testing Video
This section contains a video demonstrating the reaction of a selection of the models during
validation testing using the 2 x 2 indicator light array included in the system. The models
selected were NN-24, LGR, RF and Thresh 2 to get a broad view of the different model
types under test.

Video 4. Video demonstration of real-time detection performance for selected models

during validation testing

Artificial Neural Network Deployment Using Tinn
The deployment of most of the models studied in this work through the use of the MicroML-
Gen python package was very straightforward, however artificial neural network models are
not supported by this software at this time. The Tinn library, a dependency free C99 li-
brary which implements a single layer ANN[5], was selected due to its incredibly compact
implementation of a single layer artificial neural network, making it ideal for deployment
on embedded systems. However, the library only allowed for a single bias for the entire
hidden layer, so it had to be modified to match the default ANN implementation used in
Scikit-learn package used to train the networks which includes a bias for each node in the
hidden layer. This modified version of the library files can be found here. To initialize a
model from trained parameters Tinn requires they be grouped into two flat arrays with the
following structure:

N Nweights = {F (wT

h ), F (wT

o )} ∈ R1x(n∗m+m∗p)

N Nbias = {Bh, Bo} ∈ R1xm+p

Here wh is the n x m matrix containing the weights between the input and the hidden
layers, where ncorresponds to the number of inputs and m to the number of nodes in the
hidden layer; wo is the m x p matrix containing the weights between the hidden and output
layers with p corresponding to the number of output nodes; Bh and Bo are vectors containing
the biases for the hidden and output layers respectively; and F (M ) is an operator defining
the row-wise flattening of the input matrix M :

F : ∀M ∈nxm(cid:55)→ [row1, row2, . . . , rown] ∈ R1+n∗m

The model parameters are extracted from the trained model and included in the arduino
program as float arrays, though they could be saved to a text file on an SD card and accessed

105

via a card reader module if memory size becomes an issue. After the parameters are declared,
an empty instance of a NN with the appropriate dimensions is created using Tinn’s xtbuild()
method and the values are loaded into the model using for loops in the setup portion of the
program. To use the ANN for inference in the main loop, the input is collected and processed
as desired and fed into Tinn’s forward propagation function xtpredict() alongside the ANN
model. The ANN models used in the paper only used a single output node, so the class label
was instead determined by comparison to a threshold (0.5). More generally, however, for
classification tasks with three or more classes, the output will be a probability array from
which the predicted class can be extracted using an argmax function. Figure A-28 shows a
general implementation of the ANN deployment on an Arduino system.

Figure A-28 Rough code snippet showing the deployment of a single layer ANN model on
an Arduino microcontroller platform for the general case.

106