With rapid technological breakthroughs, the role of non-destructive evaluation (NDE) has shifted from assessing structural integrity to building complex systems with reliable defect classification and decision making capabilities. Widespread use of NDE in several industries, such as aviation, nuclear, construction and automotive, have also resulted in vastly increasing the amount of NDE information. In dealing with large volumes of data, human analysis besides being time-consuming is often inconsistent and hence there is a demand for automated signal classification (ASC) systems for accurate and consistent signal interpretation. A typical ASC system processes NDE signals, extracts appropriate features and classifies the signal categories based on signal features. Despite striking benefits of ASC systems, classification results are often affected due to inherent ambiguity of non-discriminative features, inadequate training samples or noisy measurements. As a result, uncertainty quantification in defect classification has become a critical task in NDE applications where the performance of part or structure depends on the reliability of the automated signal classification system. A reliability measure that accounts for the uncertainties in the system can help in monitoring its performance and automatically flagging indications where operator intervention is required. In addition to diagnosis, i.e., reliable characterization of current health status of industrial components, damage prognosis or prediction of their remaining-useful-life (RUL) is another essential aspect of NDE. Accurate health prognosis ensures system reliability and aids in estimating residual serviceability of a component which in turn reduces repair or replacement costs. Moreover, combining information from multiple sensors in multi-modal NDE systems can effectively improve damage growth modeling and prediction of system’s RUL. This dissertation presents three major contributions to the field of NDE diagnosis and prognosis:1. Sources of uncertainty in NDE classification have been quantified in a statistical framework to develop a confidence metric (CM) associated with ASC system output. By bootstrapping and weighting Bayes posterior probability with estimated noise distribution, effect of noise in NDE measurements is embedded into the proposed confidence metric. The effectiveness of the confidence assessment method is demonstrated on experimental data from eddy current inspection of steam generator tubes. Further, the benefit of confidence metric in improving classification performance is explored using a confidence-rated-classification technique.2. A particle filtering (PF) framework is developed for prediction of impact damage propagation in composite materials which utilizes both physical model based on Paris' law and inspection data obtained from NDE system.3. A joint likelihood updating technique is proposed in existing PF algorithm which enables optimization of damage model parameters at every time step by discarding noisy or biased measurements from multiple sources. The approach is then extended to multi-modal NDE data. Prognosis results on a composite specimen subjected to fatigue testing and inspected using two NDE modalities, validate the benefit of proposed approach over single-sensor prognosis. An additional advantage of the multi-sensor prediction in reduction of particle count within the PF algorithm is demonstrated, thereby reducing the total computation time and resources. Overall, a reliability metric and prognosis methodology is discussed for a multi-sensor system that can be extended to multiple applications.
Read
