A Multi-Headed CNN Framework for Lamb Wave-Based Damage Detection in a Thin Aluminum Plate

Abstract

In this research, a deep learning framework of 1D-Convolutional neural networks (CNNs) is utilized to recognize and auto-extract damage related features within the fragments of raw 1-D Lamb wave data. To analyze the performance of the proposed network, a diverse database is constructed to train and test the 1D-CNN architecture. It contains Lamb waves time-series data recorded from both 24 experimental and 216 Finite Element (FE) simulation setups of 1.6 mm thin Al-5052 plate. To scan the platestructure, a pitch-catch transducers configuration is adopted. The actuator-transducer excites 3.5 cycles modulated sinusoidal tone-burst signal at three different frequencies i.e., 100 kHz, 125 kHz and 150 kHz and the corresponding responses are collected from sensor-transducer only. An anomaly is introduced in the FE model in the form of a 0.8 mm deep and 2 mm wide notch-like cavity. Whereas, the damage in the experimental setup is realized by attaching an external mass of 14 gm in the center of the PZTs. Next, a 1D-CNN architecture is designed that uses raw 1D Lamb wave signals as an input and can capture high-level damage related features from the raw Lamb wave signals. Then the CNN model is trained on 90% of the total number of samples in the database using the Adam algorithm. At a later stage, the trained 1D-CNN architecture performance is validated against the 24 unseen Lamb wave responses. Out of the 24 unseen samples, the proposed architecture decisively predicts the outcome for 23 samples. This test against the unseen experimentally generated samples proves that the proposed deep learning framework has achieved the required generalization over the real scenario.