Diagnosis and classification of railway faults using deep learning methods

The study of deep learning methods has gained great momentum following developments in computer hardware, especially in graphics processor cards. These developments have enabled machine vision-based applications to achieve more successful results and expand their areas of application. In the field of object classification, the level of human performance has been surpassed and object recognition algorithms have become capable of working in real time. As deep learning algorithms are data hungry systems, the dataset is critical for training and testing. There is no publicly available railway fault diagnosis and classification dataset on the internet. In this study, a railway fault diagnosis and classification dataset has been created by compiling the images obtained from the TCDD (Turkish Republic State Railways) railway line maintenance and monitoring works and considering the most common railway faults found in the literature. With this data set, a YOLOv4 based fault detection system has been developed. During the development of the proposed fault detection system, a semi-supervised student-teacher model was developed instead of a fully supervised learning method in which all images are manually labelled and the images are automatically labelled, unlike the applications in the literature. This method reduced the human intervention in the labelling process and the high costs, and the performance of the error detection model was gradually improved thanks to the data set extended by the pseudo-labelling method. Another study within the thesis was carried out on railway fasteners. This study analysed the effect of activation function selection on the performance of the fault detection model and compared the performance of leaky, swish, x-swish and mish functions. As a result of the optimization studies, the number of True Positives (TP) was increased and the number of False Positives (FP) was reduced compared to the method used to obtain the dataset from TCDD. In the last study, a dynamic alpha-parametric focussed loss function was developed to increase the detection rate of railway faults. This emphasises the detection of the more critical fault class in railway faults and increases the diagnostic performance in this class. Experimental results show that the proposed methods provide higher accuracy and efficiency than the existing methods.