Deep Learning on Railway Images for Optimized Maintenance

Effective maintenance of railroad infrastructure is crucial for operational efficiency and cost savings. DMA, a company specializing in optimizing maintenance solutions for railroads, captured a comprehensive dataset of concrete sleepers. Expert annotators identified and annotated all faults within this dataset, creating a foundation for training a deep learning model to automatically detect such faults. Initially, DMA developed a fault detection model using YOLOv4, but its performance fell short of expectations. This project aims to enhance fault detection capabilities by utilizing the same dataset and exploring new training methodologies. Through a thorough literature review, YOLOv8 was selected for its advanced features and potential for improved accuracy. Various improvement strategies were developed, including manipulating input image size, applying data augmentations, adjusting the model's network size, and fine-tuning the learning rate. Additional factors such as dataset coherence and the impact of background noise were also investigated. Implementing these strategies resulted in an improvement in the model's performance. Whether the improvement is slight or quite substantial is hard to say due to uncertainties regarding the reported performance of the original YOLOv4 model. The mAP50 evaluation metric was increased from 41\% in the original YOLOv4 model to 73\%, while precision and recall was increased from 72\% and 62\% in the original to 76\% and 63\% in the newly trained version. Further recommendations for improvement include a deeper investigation into dataset quality and additional data augmentation techniques. Future work suggestions include training a separate model to detect concrete, thereby creating a cropped dataset with reduced background noise, and leveraging spatial and temporal data to make statistical inferences for identifying hidden faults.