تشخيص عيوب خط ريل مبتني بر شبكه يادگيري عميق

Neglecting the defects and damages of railway tracks can cause lower safety, higher costs, more repair time, and higher operation time of the railway line. Due to the frequent use of railway lines, even small defects can lead to serious accidents. Therefore, rail lines require regular inspection and maintenance. To achieve this goal, inspection and identification of defects such as breakage, cracks, other damages of rails and line components, especially those that may lead to train derailment, require a system that can detect defects in the shortest possible time. The main aim of this research is to present a rapid detection method of railway track defects based on image processing. For this purpose, by taking images of the railway line with a mobile phone camera, image processing, and training an artificial intelligence system, rail defects are extracted. In this study, two versions of the YOLOv8 model, namely n and s, are used. In this regard, 1385 data on rail and 196 data on sleepers are taken from three railway stations in Isfahan, including Silo, Iranko and Depo stations. The understudy lines have a distance of about 8 kilometers. In this study, CVAT, Python, and Colab software are used for image processing, programming, and function execution, respectively. Image processing includes noise reduction, image resizing, and light adjustment in images containing damage. After image processing and model training, a number of new images are given to it for model eva‎luation. Note that the model was not familiar with these new images. According to the results, the execution time of each image in the YOLOv8n model was about 65 milliseconds, and the execution time of each image in the YOLOv8s model was about 120 milliseconds. Comparing the execution time of both models, it is obvious that the execution time of the YOLOv8s model is approximately 1.84 times the YOLOv8n model. In the training section, the correct detection of the YOLOv8n model is 53% and the correct detection of the YOLOv8s model is 42%. In the eva‎luation section, the correct detection of the YOLOv8n model is 87% and the correct detection of the YOLOv8s model is 86%.

5

75584