Lightweight Multisensor Information Fusion Framework for Robust Monorail Transit Beam-Joint Impact Detection

Beam–joint discontinuities in straddle-type monorail tracks induce periodic vertical impacts that degrade ride comfort and long-term safety. This paper proposes a lightweight multi-sensor fusion framework that performs real-time beam–joint detection and impactearly warningby combining forward-view images with bogie speed. We develop an edge-deployable model, YOLOBJ, obtained by simplifying the YOLOv10 head to a single scale, streamlining the neck with parallel feature fusion, and introducing a Multi-Receptive-Field Fusion Layer. A constrained evolutionary neural architecture search (LEANAS) further optimizes channel widths. The final network contains 0.33M parameters, runs at an average 3.2ms per frame on an NVIDIA Jetson AGX Orin, and achieves mAP50–95= 0.785 on real-vehicle data. Field experiments with operating trains demonstrate that the proposed image+speed fusion provides predictive perception and outperforms a single-acceleration threshold baseline: recall increases by 8.9%, F1 by 5.3%, precision by 1.8%, and accuracy by 4.7%. The framework offers a practical solution for on-board deployment, supporting improved ride quality and enabling intelligent damping/control strategies for monorail systems.