A TTDCNN-based condition monitoring and crack depth severity assessment of railway wheel axle defect

Monitoring railway infrastructure is vital to prevent accidents from wheel axle cracks. Existing works largely neglected severity estimation based on crack depth. Thus, this paper introduces Tilt and Temporal Functional Deep Convolutional Neural Network (TTDCNN) for early fatigue crack prediction, combined with Weighted Cotangent Membership Functional-based Fuzzy Inference System (WCMFFIS) for severity assessment. Wheel and axle images are pre-processed via resizing, Bernoulli Distributed Contrast Limited Adaptive Histogram Equalisation (BDCLAHE)-based enhancement, and Generative Adversarial Network (GAN)-based augmentation. Then, the Red, Green, and Blue (RGB) images are converted into Hue Saturation Value (HSV) images, extracting V-components. Then, background removal, feature extraction, and feature selection using the Lyrebird Penalty Weighted Optimisation Algorithm (LPWOA) are done. A bubble map is then constructed. Simultaneously, the Fibre Bragg Grating (FBG) sensor signals are collected, pre-processed, and feature-selected to construct 3D plots. Further, TTDCNN detects axle defects, while localisation is done using Elastic Net Regularisation-based You Only Look Once version 7 (ENR-YOLOv7). Finally, severity levels are estimated using WCMFFIS. Major contribution lies in multi-modal deep learning through the fusion of image and sensor data, enabling accurate detection and severity assessment. The proposed work has detection accuracies of 99.86% for wheels and 99.65% for axles.