EMAT-Based Crack Detection in Railway Tracks Using Multi-Domain Signal Processing and Scalogram-Driven Deep Learning

This paper presents a novel crack detection approach in railroads using electromagnetic acoustic transducers (EMATs) that can be integrated with multi-domain signal processing techniques and scalogram-driven deep learning approach. In the study nine different scenarios across three critical sections of the railway track were investigated. Several useful signals processing techniques, including time-domain, frequency-domain, Power Spectrum, Periodogram, Welch Method, short-time Fourier transform (STFT), and wavelet transform, are implemented to evaluate the data acquired through EMAT sensors. Wavelet transformations are applied to the proposed segments to generate scalogram images, which are used as an input in deep learning model training. When results were compared to conventional machine learning classifiers, the deep learning model performs better, exhibiting higher accuracy in identifying different types of cracks from scalogram images. The results demonstrate that EMAT-based fracture identification, advanced signal processing, and deep learning can greatly enhance railway track inspection and safety, even though the system currently processes data in batches rather than in real time. Future work will focus on real-time data acquisition and further optimization of the deep learning architecture.