Delamination detection in aerospace composite panels using convolutional autoencoders

Modern aerospace structures demand lightweight design procedures and require scheduled maintenance intervals. Supervised deep learning strategies can allow reliable damage detection provided a large amount of data is available to train. These learning algorithms may face problems in the absence of possible damage scenarios in the training dataset. This class imbalance problem in supervised deep learning may curtail the learning process and can possess issues related to generalization on unseen examples. On the other hand, unsupervised deep learning algorithms like autoencoders can handle such situations in the absence of labeled data. In this study, an aerospace composite panel is interrogated with a circular array of piezoelectric transducers using ultrasonic guided waves in a round-robin fashion. The time-series signals are collected for both the healthy and unhealthy state of the structure and transformed into a time-frequency dataset using continuous wavelet transformation. A convolutional autoencoder algorithm trained on healthy signals is used to identify anomalies in the form of delamination in the structure. The proposed methodology can successfully identify delamination in the structure with good accuracy.