Few-Shot Mechanical Fault Diagnosis of High-Voltage Circuit Breakers Using Prior Knowledge-Constrained Wavelet-Based Multi-Frequency Feature Fusion
ID：39
Submission ID：42 View Protection：ATTENDEE
Updated Time：2025-10-11 22:28:54 Hits：63
Poster Presentation
Start Time：2025-11-09 09:04 （Asia/Shanghai）
Duration：1min
Session：[P] Poster presentation » [P4] 4.High voltage and insulation technology
No files
Abstract
Data-driven models for mechanical fault diagnosis of high-voltage circuit breakers (HVCBs) can extract health information from monitoring data. However, in practical applications, the scarcity of fault samples and severe noise interference significantly limit their feature representation ability and generalization performance. Moreover, most existing data-driven approaches overlook the underlying physical mechanisms, lacking fault-mechanism constraints, which reduces their interpretability and credibility. To address these issues, this paper proposes a few-shot fault diagnosis framework for HVCBs based on prior knowledge-constrained wavelet-based multi-frequency feature fusion. First, prior knowledge of HVCB mechanical fault diagnosis is established through excitation–response mechanism modeling, signal feature extraction rules, and simulation analysis. Second, a wavelet-based multi-frequency feature fusion network is introduced, which not only optimizes the basic features extracted by the backbone network but also incorporates frequency-domain features from different scales via wavelet transforms, thereby enhancing subtle feature representation. Third, prior knowledge constraints derived from physical processes and mechanism analysis are embedded into the fusion network, complementing the automatically extracted features of deep learning and mitigating the limitations of purely “black-box” data-driven models. Finally, a joint loss optimization function is employed to alleviate class imbalance, and metric learning is incorporated to enable effective few-shot classification. Experimental results demonstrate that the proposed method achieves a diagnostic accuracy of 98.67% and maintains high stability and generalization under conditions of noise interference and sample imbalance, highlighting its strong potential for engineering applications.
Keywords
high-voltage circuit breaker (HVCB); fault diagnosis; prior knowledge; wavelet-based multi-frequency feature fusion; physics-guided integration
Speaker
Comment submit