28 / 2025-03-26 10:10:02

Improved RT-DETR Model for Automated ROI Localization in Railway Locomotive Sanding Device Monitoring

Locomotive Sanding Device Detection, RT-DETR, MobileNetV4, Feature Enhancement, Wheel-Rail Adhesion Coefficient

During the process of braking the locomotive to a halt for maintenance, the sanding device effectively enhances braking efficiency by improving the wheel-rail adhesion coefficient, ensuring maintenance efficiency while improving operational safety. Accurate detection of the locomotive sanding device is critical for monitoring its sanding state. Aiming at the challenges of model redundancy and insufficient feature extraction in existing detection methods under resource-constrained scenarios, this paper proposes an improved RT-DETR (Real-Time Detection Transformer)-based detection method for locomotive sanding devices. The method replaces the original ResNet-18 with MobileNetV4 as the backbone network, significantly reducing model parameters while maintaining performance. Additionally, a Local Feature Extraction (LFE) module is introduced to optimize the RepC3 (Reparameterized Convolution Block 3) module, enhancing the network’s representational capability for fine-grained features of the sanding device. Experimental results demonstrate that the improved RT-DETR model achieves a 1.9% increase in mean average precision (mAP50) while reducing model parameters by 43.7% and computational complexity (GFLOPs) by 32.3%, providing an industrial-grade solution for intelligent operation and maintenance in rail transit.