LAMB WAVE-BASED STRUCTURAL HEALTH MONITORING USING DEEP LEARNING TECHNIQUES


Abstract:



Lamb waves-based structural health monitoring (SHM) has emerged as one of the most feasible options for implementing affordable onboard SHM systems. A data-driven machine learning (ML) model can analyze the complex Lamb wave responses to diagnose the health of the structure. The research aims to discuss the various challenges associated with data-driven SHM and apply different ML and deep learning (DL) architectures to monitor the health of thin-walled structures. The study is conducted in five phases. Initially, various conventional ML models and artificial neural networks (ANN) are utilized to detect damage in a thin aluminium plate using an automated feature extraction tool that analyzes Lamb wave responses. The effectiveness of these models largely depends on the robustness of the features used for their training. In the second phase, a one-dimensional convolutional neural network (1D-CNN) has been applied for diagnosing damage in carbon fibre reinforced polymer (CFRP) plate using diagnostic Lamb wave data obtained from an open source database, namely, openguidedwaves (OGW). The use of CNN eliminates the need for explicit feature extraction and signal processing. Most of the DL techniques are data-demanding in nature and obtaining a comprehensive dataset in domains like aerospace is challenging due to restricted data access, sensor deployment difficulties, high costs, and rare events. The third phase of the study utilizes a semi-supervised generative adversarial network (SGAN), which can be trained on smaller datasets, for damage detection in CFRP plates using the OGW dataset. The proposed SGAN model demonstrated better generalization compared to the CNN and the ResNet autoencoder-based transfer learning (TL) model. The fourth phase applies the transformer model for detecting the presence and extent of disbonding in a stiffened aluminium plate by analyzing Lamb wave responses obtained through an in-house experimental setup. The transformer outperformed the CNN model when tested on a noisy dataset. Finally, two different CNN models are applied for disbond detection. First, a CNN-classifier (CNN-C) classifies the health of the stiffened plate in a binary classification problem, and then a CNN-regressor (CNN-R) estimates the width of disbonding in a regression problem. In summary, this thesis demonstrates the application of diverse ML and DL techniques for SHM across different structures, while tackling some of the key challenges such as data scarcity, complexity in wave interactions, and the need for robust damage detection.



Keywords: Structural health monitoring, Lamb waves, signal processing, piezoelectric transducers, CFRP composites, stiffened panels, ANN, 1D-CNN, SGAN, Transformer.