A Deep Learning Approach for Inspection of Locomotive Undercarriage Components

Regular inspection and preventive maintenance of railway vehicles are essential for ensuring the safety, reliability, and operational availability of railway systems. However, visual inspection of locomotive undercarriage components is still predominantly performed manually, which is time-consuming, labor-intensive, and susceptible to subjective assessment errors. This paper presents a deep learning-based framework for automated visual inspection of locomotive brake system components under realistic operating conditions. The proposed methodology integrates YOLOv4-tiny-based object detection with convolutional neural network (CNN) classification models to localize brake components and identify their operational states from digital images. The classification task was formulated as a three-class problem, distinguishing between two non-functional states (missing brake shoe and key, and missing key) and the correct operational condition with both components present. Classification performance was assessed using accuracy, macro precision, macro recall, macro F1-score, Confusion matrices, and Receiver Operating Characteristic (ROC) curves with the corresponding Area Under the Curve (AUC-ROC) values. The YOLOv4-tiny detector achieved an AP@0.5 value of 92.55%, enabling accurate localization of locomotive brake system components prior to classification. The second, more rigorous metric, mAP@[0.50:0.95], represents the mean precision calculated over the range of IoU thresholds from 0.50 to 0.95 with a step of 0.05, and it is 49.36%. Among the evaluated classification models, ResNet-18 demonstrated the optimal classification performance, achieving 99.69% accuracy and a macro F1-score of 99.66% with a processing speed of 98.1940 FPS. When integrated into the complete detection-classification pipeline, the system achieved an overall accuracy of 93.33% on the independent test set. Classification errors involved samples with a missing brake shoe key being identified as functional, revealing a safety-critical limitation. The results indicate promising potential for automated visual inspection of locomotive brake components and maintenance support, with further improvement and validation required to address these errors.