Detection and Severity Evaluation of Combined Rail Defects Using Deep Learning

Various techniques have been developed to detect railway defects. One of the popular is machine learning. This unprecedented study applies deep learning, which a branch learning techniques, and evaluate severity rail combined The defects in are settlement dipped joint. Features used axle box accelerations simulated using verified rolling stock dynamic behavior simulation called D-Track. A total 1650 simulations run generate numerical data. Deep neural network (DNN), convolutional (CNN), recurrent (RNN). Simulated data two ways: simplified raw Simplified develop DNN model, while CNN RNN model. For data, features extracted from weight stock, speed three peak bottom wheels stock. In total, there 14 as for developing time-domain directly models without processing extraction. Hyperparameter tuning performed ensure that performance each model optimized. Grid search performing hyperparameter tuning. To defects, proposes approaches. first approach uses one joint, second joint separately. results show both approaches provide same accuracy 99%, so good enough classification regression concepts. Classification by categorizing into light, medium, severe classes, estimate size From study, suitable evaluating with an 84% mean absolute error (MAE) 1.25 mm, 99% 1.58 mm.