Classification of rail defect based on B-type display image using deep learning method

• Main Author: Li, Jie
• Format: Thesis
• Language: English
• Published: 2023
• Subjects: T Technology (General); TA Engineering (General). Civil engineering (General)
• Online Access: http://umpir.ump.edu.my/id/eprint/38497/1/ir.Classification%20of%20rail%20defect%20based%20on%20B-type%20display%20image%20using%20deep%20learning%20method.pdf; http://umpir.ump.edu.my/id/eprint/38497/

The rail defect detection is the main method to ensure that the railway transportation is safe. The availability of rail defect information enables the railway departments to determine the integrity of the steel rail and provide suitable plans for railway operation and maintenance. However, the current rail defect detection still relies on the traditional method which require high manpower intensity and time consuming. The high manpower at the current state is unable to cater for the growing need of the railway industries. Furthermore, the traditional method is prone to errors and mistake which reduces the accuracy of the defect detection process. Therefore, this study aims to propose an automated recognition method based on machine learning and image processing to provide more efficient defect detection process while reducing the need of manpower. To achieve that aim, the objectives are to: (1) To classify the steel rail defect by using manpower; (2) To develop deep learning models to classify steel rail defect based on B-type display image; and (3) To optimize deep learning models with different variations of epoch. In phase 1, a total of 6000 rail defect images has been collected from China Railway Hohhot Railway Department. The defects were classified and identified. In phase 2, a newly developed model ResNet50 has been developed for steel rail defect identification and classification. This study uses 5000 steel rail defect images as training data to train ResNet50 model, and then using 1000 steel rail images as testing data to validate model structure. In phase 3, the newly developed ResNet50 are optimized by varying the parameter values of the model framework, 14 final data analysis results were finally obtained. The analysis of the fit and convergence of data results shows that the ResNet50 model can obtain optimal results at Epoch11. This study found that the overall accuracy of the proposed ResNet50 model was 100% in the test dataset and the detection time of a single defect image was 156 ms/ image, while the remaining three deep learning GoogleNet, VGGNet and AlexNet methods were <95%. The comparative results show that the proposed ResNet50 model has the potential to be applied to the automatic identification and classification of large-scale rail defects.