Multi-modal sensing and pulse sequence analysis for single-source and dual-source partial discharge diagnosis in transformers under complex operating conditions
Transformer partial discharge (PD) diagnosis may simultaneously face narrowband interference under undersampling conditions, limited fault samples, class imbalance, and multi-source signal mixing. To address these issues, this paper proposes a multi-modal pulse-sequence-based diagnostic framework using synchronized Optical, ultra-high-frequency (UHF), and high-frequency current transformer (HFCT) measurements, and experiments are conducted on a laboratory platform with five typical PD defect models of oil-immersed transformers. For front-end signal processing, a spectral dilation and linear trend replacement (SDLTR) method is proposed to suppress narrowband interference in HFCT signals while preserving the original pulse timing and amplitude characteristics. On this basis, conventional single-sensor pulse sequence analysis (PSA) is extended to an adaptive tri-modal PSA fusion scheme for single-source PD classification. By constructing the temporal union of synchronized Optical, UHF, and HFCT pulse streams and using expert-weighted decision fusion, the proposed method exploits cross-modal complementarity and enlarges the effective sample set. Under class-imbalanced conditions, the four-pulse-based PSA6 fusion scheme achieves an accuracy of 95.47% and a Macro-F1 of 95.17%. For dual-source PD mixtures, an adaptive cascaded decoupling framework (ACDF) is further proposed by combining class-level precision-weighted fusion, an adaptive confidence boundary, and two-stage dominant-source stripping based on PSA6 and PSA4. The proposed framework produces zero false decisions in single-source verification and correctly identifies both PD sources in all ten dual-source combinations. These results demonstrate that the proposed framework provides an effective and practical solution for transformer PD diagnosis under complex operating conditions.