Abstract
Gas-insulated switchgear (GIS) insulation defect diagnosis faces persistent challenges due to heterogeneous multisensor signals, the lack of structured fusion across sensor types, and severe catastrophic forgetting under source-free streaming conditions. To address these issues, this article proposes a two-level fusion multisensor network with incremental learning (TFMNIL) for on-site diagnosis. Unlike conventional one-stage fusion strategies, the proposed framework separates multisensor fusion into channel-level fusion of homogeneous sensors to preserve consistent physical characteristics, and feature-level fusion of heterogeneous sensors to reduce cross-modal semantic discrepancies and redundancy. To support continual adaptation under evolving operating conditions, an incremental domain adaptation strategy is further developed. It integrates dynamic prototype updating to compensate for feature drift with classifier-weight–anchored cross-domain contrastive alignment, enabling stable class-level supervision without access to source data. Experiments on large-scale real-world multisensor GIS datasets show that TFMNIL achieves a diagnostic accuracy of 95.86% and maintains strong adaptability to newly emerging fault types in practical on-site scenarios.
Funding Information
  • National Key Research and Development Program of China (2022YFB2403700)

This publication has 36 references indexed in Scilit: