Insulation condition assessment of 10kV cable joints based on multi-dimensional feature extraction and D-S decision fusion

Abstract

A two-tier cascaded assessment framework based on multi-dimensional feature extraction and Dempster-Shafer (D-S) decision fusion is proposed to address the physical blind spots inherent in single-parameter monitoring and the boundary ambiguity caused by high-noise environments during the insulation deterioration of 10 kV cable joints. At the feature level, a hybrid neural network incorporating the physical prior of a two-node Lumped Parameter Thermal Network (LPTN) is constructed. Partial Discharge (PD) signals are mapped into two-dimensional topological matrices utilizing Discrete Wavelet Transform (DWT) and Phase-Resolved Partial Discharge (PRPD) techniques, while temperaturecurrent sequences are synchronized via a sliding window mechanism. Subsequently, a dual-branch architecture comprising a modified single-channel ResNet-18 and a 1D-CNN-LSTM is utilized to achieve the dimensionality reduction and spatio-temporal alignment of microsecond-level PD spatial topologies and hour-level electro-thermal inertia characteristics. At the decision level, to resolve the conflict among multi-source sensing information, an improved D-S evidence theory based on a dynamic conflict coefficient K and a penalty factor β is introduced. This mechanism penalizes and distributes high-conflict beliefs equiprobably into independent state subspaces, thereby eliminating falsepositive misjudgments triggered by single-sensor node anomalies. Experimental validation based on 24,000 heterogeneous data pairs demonstrates that the proposed method achieves an overall assessment accuracy of 96.8% under strong perturbation conditions. The results indicate excellent diagnostic robustness and the potential for localized deployment in edge computing gateways.