Sensor integrity assessment in guided wave SHM using convolutional neural networks

Abstract Guided wave (GW) based structural health monitoring (SHM) systems are extensively utilized for real-time inspection of structures. The traditional SHM system assumes ideal bonding between the sensors and the host structure. In practice, however, complete bonding is rarely achieved due to the environmental and operational conditions, often leading to partial sensor debonding. Debonding may significantly impact the signal interpretation, resulting in the false damage identification or reduced accuracy. Existing methodologies, like maximum amplitude spectra (MAS) analysis and electro-mechanical impedance techniques, are limited by their dependence on the sensor placement geometry and are unable to quantify the degree of debonding. This article presents a data-driven, automated methodology employing a one-dimensional convolutional neural network (1D-CNN) to the detect sensor debonding inside a square sensor network. The model is trained on experimentally obtained GW responses from an aluminum plate containing a rectangular sensor network with varying degrees of debonding. The CNN approach detects the debonding, identifies the impacted sensor, and predicts the degree of debonding. The model achieves about 100% accuracy and exhibits robust generalization when evaluated using the stratified cross validation method. A comparative assessment indicates that the CNN model outperforms the conventional MAS-based techniques. Explainable AI methodologies like layer-wise relevance propagation are employed to clarify the CNN predictions, indicating that significant time-domain variables linked to the wave propagation modes influence the model’s decision-making process. The suggested 1D-CNN-based approach ensures the sensor integrity before assessing the structural damage, serving as a reliable pre-screening tool that offers a robust framework suitable for a complex geometry and the diverse sensor network configurations.