A Novel Hybrid Deep Autoencoder Approach for Unsupervised Loosening Detection in Bolted Lap Joints
Health monitoring of bolted joints is crucial in guaranteeing structural integrity and safe operation in many engineering applications by enabling quick and accurate detection of bolt loosening, which is among the most common failure modes. Guided wave-based Structural Health Monitoring (GW-based SHM) systems are generally preferred for thin, plate-like structures given that they can identify minute structural issues and survey extensive areas effectively with a minimal number of transducers. In this work, experiments are carried out using piezoelectric transducers mounted on an aluminium lap joint with two bolts to evaluate the Lamb wave propagation characteristics across the joint under various stages of loosening. Loosening is simulated by varying torque values on the bolts from 10 Nm down to 3 Nm, with a step size of 0.5 Nm, using a torque wrench. Three additional torque conditions, namely, a hand-tightened state, a state where the bolt is present but zero torque is applied, and a state where the bolt is completely absent, are also evaluated. Furthermore, seventeen different excitation frequencies spanning a wide range from 52 kHz to 260 kHz are employed for generalization. Investigation of the signals collected from the experiments reveals a clear reduction in the signal amplitude as the bolts are progressively loosened. This result is consistent with the reduction in the amount of wave energy transmitted across the joint under loosening due to the reduction in the contact area between the plates in the joint overlap region. The central objective of this study is the development and evaluation of a novel unsupervised deep learning framework tailored for loosening detection, which centers on a comparative study of two hybrid autoencoder architectures: the CNN Encoder-GRU Decoder (C-G AE) and the GRU Encoder-CNN Decoder (G-C AE). Convolutional Neural Networks (CNNs) perform hierarchical feature extraction by capturing local spatial patterns and short-term dependencies. On the other hand, Gated Recurrent Units (GRUs) model the long-term dependencies and temporal dynamics. Combining both leverages the capability to extract localized features and their temporal evolution. This study enables a comparative evaluation of how effective the spatial-first versus temporal-first encoding strategies are in terms of latent representation of the time-series data, reconstruction fidelity, and the anomaly detection performance. Additionally, we will systematically investigate the influence of the ordering of the CNN and GRU components, as well as the impact of varying levels of noise and the volume of training data, on the overall anomaly detection performance. The models will be trained exclusively on minimally processed time-series signals representing the torqued conditions by minimizing the reconstruction error. The unsupervised learning paradigm helps in avoiding the enormous task of labeling all the possible damage scenarios. The performance of the trained models will then be assessed using unseen signals representing loosened conditions. By rigorously comparing the models' capacity to correctly detect anomalous signals, this work aims to identify the most efficient and reliable hybrid deep learning architecture for the monitoring of joint integrity.