Detecting Far-Side Corrosion of Multi-layer Structure on Aircraft with Time-Series Signal and Explainable Deep Learning Approach

Electromagnetic testing (ET) stands out as a highly efficient technique within nondestructive testing (NDT) practices for evaluating the far-side corrosion of aircraft structures. Typically, concealed corrosions around the rivet pose challenges in the detection. In this study, we leverage Deep Learning approach to solve this problem. Our key innovation involves treating these concerns as time-series data rather than 2D images as the previous researches. Unlike conventional approaches, we tackle redundant activation map issues, adapting the GhostModule. Furthermore, we provide comprehensive insights into our findings, employing Explainable Artificial Intelligence (XAI) techniques – gradient-based method for a detailed and transparent evaluation of the results obtained. Achieving a commendable average classification accuracy of 93% but maintains a lightweight profile, utilizing only 37K parameters and 3.67 miliseconds of inference time on an STM32 Microcontroller (NUCLEO-H743ZI2). This breakthrough strikes a crucial balance between accuracy and efficiency, showcasing promise for detecting hidden corrosion volumes ranging from 2.8–195.4 mm3.