Distribution Network Fault Screening Method Based on Multimodal Information Sensing and Deep Feature Fusion

Abstract

: As distribution network sensing terminals expand from a single electrical measurement point to multi-source heterogeneous data such as infrared thermal images and operation and maintenance logs, the massive information provides rich fault diagnosis clues for intelligent operation and maintenance. However, most existing research still focuses on unimodal transient waveforms, and accuracy and recall drop significantly when facing hidden faults such as high-resistance grounding, leakage, etc., making it difficult to meet the demand for rapid screening under complex working conditions. For this reason, this paper adopts the method of “multimodal information perception + deep feature fusion” to model waveform, infrared (IR), and text tri-modal data uniformly, to enable efficient fault screening and assessment in distribution networks. First, this paper constructs an accurate time-synchronized multimodal acquisition system and proposes the corresponding hierarchical preprocessing and coding strategies; then, it designs a fault screening framework with convolutional cross-attention and cross-modal dynamic gating to achieve heterogeneous feature alignment and fusion; finally, it introduces a multitasking filter in the unified embedding space, which synchronizes the output of fault triggering, type identification and severity assessment. The experimental results show that the proposed method achieves an overall accuracy of 97.8% on the simulation-measurement hybrid dataset, which is 5.7 percentage points higher than the unimodal optimal baseline. The recall rates of high-resistance grounding and leakage faults are increased to 96.3% and 95.1%, respectively, and an average of 4.1 min of warning is realized in the field pilot of a 10 kV urban network, which verifies the value of its application in real distribution networks.