EXPLAINABLE MACHINE LEARNING FRAMEWORK FOR GUIDED WAVES SIGNAL RECONSTRUCTION AND STRUCTURAL HEALTH MONITORING UNDER VARYING OPERATING AND ENVIRONMENTAL STATES

Abstract

Lamb Wave (LW) signal has been widely used as a technique for damage identification and localization in Structural Health Monitoring (SHM) due to its sensitivity to varying types of state changes. Analyzing and constructing the guided wave signals then become a critical step in damage detection and assessment. Researchers have discovered features in both temporal and frequency domains for signal description and reconstruction. Yet the related features are challenging to be manually developed as the stochastic acoustic signals captured by sensors can be complex and are determined by various factors such as complex boundary conditions and material properties. Recently, neural network has exhibited the capability for time series reconstruction yet lacking of interpretability. In this study, a convolutional autoencoder (CAE) network has been developed to compress the information of the collected signals along with parameters such as damage levels and external loads into time-invariant and time-variant latent spaces at the bottleneck layer which can be easier to analyze and more efficiently used for state estimation and signal reconstruction. The power of estimating signal and its corresponding conditions has been examined by combining a feed forward neural network (FFNN) with the encoder or decoder extracted from the CAE network so that states of raw signals can be predicted and signals under known states can be reconstructed. The proposed framework has been applied in two test cases to verify its capability and stability in terms of different latent space types. The experiment was conducted on an Al plate under different damage states with PZTs serving as actuators and receivers. It is shown that the state parameters can be estimated with high accuracy and the signals can be generated with low error and thus alleviate the requirements of time-consuming experiments.