Digital Validation and Interpretability Study of Measured Assets Based on Multi-Task Regional Proposal Networks

Verifying the consistency between digitized files and physical objects of power metering assets is a key link in the construction of smart grid. In this paper, an innovative Multi-Task regional proposal network (MT-RPN) framework is proposed to address the problems of inefficiency of traditional single-task detection models, insufficient robustness in complex scenarios and lack of decision interpretability. The model takes ResNet-101 as the backbone network, and branches out three collaborative task heads through the feature pyramid architecture: appearance verification, meter position verification and device verification. It innovatively introduces the coordinate attention mechanism to address the challenge of small target detection, and designs a dynamic task balancing strategy based on uncertainty. In terms of interpretability, it incorporates the fusion of feature visualization and knowledge graph-driven natural language report generation technologies. Experiments on a dataset of 4 million on-site metering box images show that the model's integrated detection accuracy ([email protected]) reaches 92.7%, an improvement of 9.2 percentage points over the single-task baseline; and the efficiency of a single verification is improved by 32.5% after deploying it to a power grid.