An Automatic Multi-Scale Feature Learner and a Transformer Algorithm for Fault Detection

Partial discharge (PD) in covered conductors (CCs) indicates the risks of latent faults and significant insulation degradation. Precisely identifying PD patterns is vital for maintaining electrical systems. A framework for recognizing PD patterns in overhead CCs based on an automatic multi-scale feature learning network and a Transformer is introduced in this paper. First, the method effectively removes background noise via the periodic settings of a multi-seasonal time series decomposition algorithm. An automatic feature multi-scale learning network is then constructed to learn signal features, aiming to minimize the degree of manual intervention. It enhances time series data on the basis of three-phase signal features to address class imbalance problems. An innovative global multichannel pattern recognition framework utilizing a Transformer is designed, featuring positional encoders to identify intra-phase and inter-phase feature correlations and a dynamic gating mechanism for capturing complex data patterns. In experimental validations, the proposed algorithm achieves a detection accuracy of 98.6% and a specificity of 99.2%, representing an superior performance in this field. This research provides an accurate and highly generalizable solution for PD detection, offering solid theoretical support for the digital operations and maintenance of power transmission and distribution equipment.