<p>During train operations, the frequent occurrence of pantograph-catenary arcs, caused by changes in the external operating environment, presents serious safety risks. Efficient detection of these pantograph-catenary arcs is vital for pantograph system maintenance. Traditional methods for detecting pantograph arc are time-consuming and rely heavily on manual feature design. Addressing these challenges, this study introduces a novel pantograph arcing detection model. This model overcomes issues of low accuracy and slow processing speeds and align text information with image features to achieve rapid diagnosis of pantograph status. Firstly, We introduce an efficient feature extraction convolution module in the ViT backbone to improve performance while reducing computational costs. Secondly, Adopt an information guided Contrastive perception feature fusion method to efficiently fuse three different hierarchical features, During this process, local features extracted from CNN interact globally with high-level semantic information from EfficientViT. Finally, By incorporating a method of novel upsampling method, allowing for the model to perform more streamlined feature transformations while maintaining high efficiency. This model achieves 92.7% on mAP and 85.3 on FPS. In comparison to the prior arc detection model, Our algorithm has significantly improved detection accuracy and significantly reduced computational cost. These results provide robust technical support for the safety monitoring of pantograph performance in high-speed trains and the deployment of edge devices.</p>