Automated debonding assessment of tile panels using convolutional neural network

Deep learning networks have been widely applied in signal processing for structural health monitoring, enabling a more comprehensive and automated analysis. However, existing research primarily focused on identifying faulty signals and the presence of defects, with few studies addressing defect quantification. Particularly, previous research on invisible defects has predominantly targeted delamination in carbon fibre-reinforced polymer and fibre-reinforced polymer-concrete, with limited attention given to the tiling systems. Although techniques have been developed for identifying debonding in such structures, research on debonding quantification has been limited due to the challenges of analysing and differentiating complex vibration signals. This study thus aims to develop an automated system for classifying the debonding size of tile panels based on their vibration responses. The proposed approach involved transforming the waveform data into scalograms, followed by augmentation to prepare the dataset for network training. An ad-hoc convolutional neural network was designed to categorise the vibration data into three classes that represent different ranges of debonding sizes. The network training involves both simulated and experimental data, achieving high accuracy in predicting simulated data while having relatively lower accuracy for experimental data. This can be attributed to the discrepancies between simulated and experimental data, as well as the challenges in identifying cases at the boundaries of two adjacent classes. The limited size of the experimental dataset may also have constrained the performance of the trained network.