Common-Specific Subspace Decomposition Based Deep Ensemble Representation for Multimode Process Monitoring in Power Systems

Transitions in power system operating conditions introduce multimode characteristics, posing significant challenges to process monitoring, such as increased false alarms and missed detections. Due to nonstationary feature distributions and strong dynamic coupling among different modes, existing methods suffer from insufficient representation of mode-related features. In addition, single models often struggle to maintain robustness under significant distribution shifts. To address these issues, this work proposes a deep common-specific feature decomposition (DCSFD) that disentangles multimode coupling dynamics through a layered ensemble representation. First, multiple common-specific decomposition blocks are cascaded to progressively suppress dominant mode-related variations and enhance subtle fault signatures. Second, a direct-variable-constrained decomposition mechanism is developed to impose low-rank constraints directly on the common latent variable, improving feature interpretability and distinguishing DCSFD from existing shared-specific decomposition methods. Theoretically, a mathematical analysis derives a lower bound for the detection index, establishing theoretical support for detection performance under noise interference. The effectiveness of the proposed method is demonstrated on both a simulation and a real-world three-phase inverter platform, where DCSFD achieves improvements of 16.62% and 31.88% in fault detection rate, respectively.

© 2026 IEEE. All rights reserved, including rights for text and data mining, and training of artificial intelligence and similar technologies. Personal use is permitted, but republication/redistribution requires IEEE permission.