Tiny Machine Learning Implementation for Guided Wave-Based Damage Localization.

Abstract

This work leverages ultrasonic guided waves (UGWs) to detect and localize damage in structures using lightweight Artificial Intelligence (AI) models. It investigates the use of machine learning (ML) to train the effects of the damage on UGWs to the model. To reduce the number of trainable parameters, a physical signal processing approach is applied to the raw data before passing the data to the model. Starting from current state of the art in algorithms used for damage detection and localization, an AI-based technique is developed and validated on an experimental benchmark dataset before tiny ML implementation on a low-cost development board. A discussion of the need for a balance between the reduction in computational resources and increasing the precision of the models is also reported. It is shown that by extracting simple features of the signal, the models required to predict the damage locations can be significantly reduced in size while still having high accuracies of over 90%. In addition, it is possible to use these predictions to construct a fairly accurate heat map indicating the likely damage locations. Finally, a convenient edge/cloud visualization of the results can be achieved by simplifying the heat map.