DEEP LEARNING APPROACH TO REAL-TIME HEALTH MONITORING FOR FATIGUE DAMAGE DETECTION AND CLASSIFICATION

Abstract

The work, reported in this thesis, focuses on the development of an automated monitoring system for detecting and classifying fatigue damage in mechanical structures. The thesis starts with a background of fracture theory, which aims to investigate the theory and practice of crack propagation in ductile materials, and it has two main objectives. The first objective is to clearly explain the notion of fracture mechanics and the fracture theory (e.g., how the crack initiates and extends on the surface of the material) by stating appropriate definitions. The second objective is to study the concept of fatigue failure and establish how to calculate the remaining life of the fatigue damaged component. Part of this thesis concerns the forecasting and detection of fatigue cracks, where they play a key role in damage mitigation of mechanical structures to enhance their service life. Ultrasonic testing (UT) has emerged as a powerful tool for detection of fatigue cracks at early stages of damage evolution. Along this line, part of the work reported in this thesis aims to improve the performance of fatigue crack forecasting and detection, based on a synergistic combination of discrete wavelet transform (DWT) and Hilbert transform (HT) of UT data. The performance of the proposed method is evaluated by comparison with the images generated from a digital microscope, which are treated as the ground truth in this study. The results of the comparison show that forthcoming fatigue cracks can be detected ahead of their appearance on the surface of specimens. The proposed method outperforms both HT and conventional DWT, when they are applied individually because the synergistic combination of DWT and HT provides a better characterization of UT signal attenuation for detection of fatigue crack damage. The methodology of feature extraction, based on DWT, HT, and their combination, is applied for detection and classification of fatigue damage in mechanical structures in the framework of neural networks (NN). These features are (i) low-frequency signal contents, (ii) high-frequency signal contents, (iii) signal envelope, and (iv) signal energy. The performance of the constructed NN model is evaluated in terms of its accuracy, precision, sensitivity, and specificity. The efficacy of the proposed methodology is determined from the area under the curve (AUC) of receiver operating characteristics, the Matthews correlation coefficient (MCC), and the F-Measure. The performance improvement of each feature in the NN model is compared with that of the original UT data. The results show that feature extraction has a significant impact on the performance of the NN model. The best feature is capable of detecting and classifying the fatigue damage with (up to) 98.5% accuracy; and MCC shows that the correlation between the target model responses and predicted responses is strongly correlated. iii Another part of the thesis focuses on building an automated health monitoring system. Based on an ensemble of time series of ultrasonic test (UT) data, the proposed procedure is found to be capable of detecting fatigue-damage (at an early stage) in mechanical structures, which is followed by an online evaluation of the associated risk. The underlying concept is built upon two neural network (NN)-based models, where the first NN model is the best feature model, while the second NN model further classifies an identified damaged structure into three classes: low-risk, medium-risk, and high-risk. The input information to the second NN model is the crack tip opening displacement (CTOD), which is computed by the first NN model via linear regression from an ensemble of optical data, acquired from the experiments. Both NN models have been trained by using scaled conjugate gradient algorithms. The results show that the second NN model classifies the risk of fatigue damage wi