Diffusion-Based Anomaly Detection for Railway Track Fault Diagnosis

Abstract

Early detection of railway rail surface defects is critical. Late or missed detection of these defects can result in loss of life and property. Monitoring the smart maintenance processes of railway rail lines, a critical infrastructure within the scope of Industry 4.0, offers a significant advantage. This study presents an innovative method for anomaly detection of potential rail surface defects on railway lines. The proposed approach is based on a diffusion model and accelerates and improves the maintenance process by detecting rail deterioration early. This method not only eliminates the need for data labeled as faulty or healthy, but also allows the model to predict fault types previously unseen using unsupervised learning on test data. Model training is completed using only healthy rail images. In the test phase, the model reconstructs the faulty images and calculates anomaly scores from pixel-level differences. Experimental studies were conducted on a publicly available rail fault dataset to evaluate the method's performance. The study also examined the distribution of anomaly scores, score histograms, and heat maps in detail.