A Masked Autoencoder-Based Novel Deep Learning Framework of Delamination Detection Using Improved HHT-Based Signal Representations

Abstract

Ensuring the longevity and operational efficiency of composite structures demands highly effective structural health monitoring (SHM) strategies, particularly for detecting delamination, a critical defect that severely compromises structural integrity. Guided ultrasonic wave (GUW) methods have proven particularly adept at identifying the presence, location, and extent of delamination. Despite these advantages, implementing conventional, supervised deep learning (DL) approaches for GUW-based SHM has been constrained by the scarcity of high-quality, labelled training data and the high computational expense of generating representative damage scenarios. To overcome these limitations, this study introduces an unsupervised DL framework for delamination detection that incorporates a masked autoencoder (MAE) architecture. Firstly, one-dimensional time-series signals from GUW experiments are converted into two-dimensional time-frequency representations using an enhanced Hilbert-Huang Transform, which is chosen for its superior time-frequency resolution and computational efficiency over wavelet-based approaches. In the subsequent phase, the MAE framework is trained exclusively on these wavefield representations corresponding to undamaged conditions, allowing it to excel at reconstructing masked segments. The model is then presented with new, unlabeled data that includes both undamaged and damaged signals. If the input data significantly diverges from the patterns learned during training, the model’s reconstruction of masked regions deteriorates, resulting in higher reconstruction error that serves as an indicator of potential damage. A benchmark of an experimental dataset from the open guided waves platform was used to validate this framework. This dataset encompasses carbon fibre-reinforced polymer plates with 28 distinct damage scenarios. The framework identifies delamination damage through root mean squared reconstruction errors and adaptive thresholds and has been found to outperform state-of-the-art DL architectures, including conventional autoencoders.