Research on Multi-source Data Fusion Edge Pattern Recognition Method Based on Lightweight Deep Network

Aiming at the problems of difficult multi-source heterogeneous data fusion, limited resources of edge terminals, and insufficient recognition accuracy of small-target faults in power inspection and equipment state perception, this paper proposes a multi-source data fusion edge pattern recognition method based on lightweight deep network. The method constructs a cloud-edge-end collaborative recognition architecture, where multi-source data collection and preprocessing are completed at the end side, real-time fusion inference is realized at the edge side, and model training, updating and complex sample verification are completed at the cloud side. In terms of model design, this paper adopts a local-global dual-branch feature extraction structure, combining differential convolution and efficient visual Transformer to enhance the expression ability of local defect details and global semantic modeling; further, a multi-dimensional coordinate collaborative attention fusion module is designed to realize the adaptive fusion of visible light, infrared, difference images, time-series signals and context information; meanwhile, a small-target enhanced multi-scale detection head is introduced to improve the recognition ability of small-scale faults such as insulator damage, strand breakage, hardware loosening and partial discharge. Experimental results show that the proposed method achieves 99.2% mAP in power defect detection tasks. The compressed model significantly improves the inference speed while maintaining 98.7% mAP, and has real-time or near-real-time deployment capability on edge platforms such as Jetson Nano. The results verify the effectiveness of the proposed method in terms of recognition accuracy, multi-source fusion robustness and edge inference efficiency.