Attention enhanced edge vision measurement for real time pantograph catenary arc monitoring

Reliable pantograph-catenary arc (PCA) monitoring is important for high-speed railway power-supply safety, but field visual measurement is challenged by small arc regions, transient and irregular morphology, intense halos, and background highlights caused by tunnels, stations, rain, snow, and strong sunlight. This paper proposes ACE-PCANet, an attention-enhanced edge vision measurement framework for real-time PCA event localization. In the proposed measurement chain, the onboard visible-light sensor provides image observations, ACE-PCANet performs measurement-oriented data processing, and the output is an arc event indication with a bounding-box location referenced to manual annotations. The framework integrates three task-oriented modifications into a one-stage detector. First, a Convolutional Block Attention Module (CBAM) reweights channel and spatial features to address high-luminance background interference while retaining arc-related features. Second, a Coordinate Attention Fusion Module (CAFM) is introduced in the neck to fuse low- and high-level features with position-aware attention, providing position-sensitive contextual information for feature fusion. Third, Efficient IoU (EIoU) improves bounding-box regression for arc targets with highly variable aspect ratios. Experiments on 5,625 field images collected under four operating scenarios show that ACE-PCANet achieves 94.1% precision, 92.7% recall, and 91.5% [email protected]. Scenario-wise tests further indicate improved robustness under strong light, tunnel, and rain/snow/fog conditions. A lightweight variant, ACE-PCANet-lite, is deployed on an NVIDIA Jetson AGX Orin using TensorRT INT8 acceleration and reaches 89.2% [email protected] and 91.19% F1 after quantization, with 68 FPS, 14.7 ms per-frame latency, and 15.2 W measured system power. The reported results characterize the data-processing accuracy, robustness, and edge-instrument feasibility of the visual measurement framework; they are not intended as a formal metrological uncertainty budget.