A unified CNN approach for guided wave-based damage detection, damage size estimation and reliability assessment demonstrated on a complex composite structure

This work presents a novel unified Convolutional Neural Network approach where broadband ultrasonic guided waves signals are processed in such a way that damage is first detected (binary classification) and then its severity assessed on continuous scale (multi-class classification) without resorting to different procedures. To test the learning approach and assess the classification procedures, a hyperparameter optimization is first carried out to determine the best data processing procedure. Then, the performance of the network is evaluated thoroughly. The results demonstrated the relationship between the model’s performance and SHM system parameters, including excitation signal, pre-processing approach and the number of paths utilized within a sparse distributed transducer network. Furthermore, the damage location is an important influence factor. In addition to that, ensemble voting is demonstrated to be the most accurate approach to achieve high reliability in damage detection and size assessment. The results show the capability of the proposed methodology (i) to detect early damage with highest possible accuracy (ii) to estimate the dimension of damage with limited error and reasonable accuracy, and (iii) to assess the reliability of the whole monitoring system through damage size estimation combined with a critical damage size approach.