An Adaptive Sliding Window-Based Online Data-Driven Diagnosis Method for Inverter Open-Circuit Fault in WRSM Drive Systems

This article presents a novel data-driven real-time open-circuit (OC) fault diagnosis strategy for three-phase inverters, integrating the adaptive sliding window (ASWIN) method with machine learning algorithms, enabling 22 state classifications. In addition to stationary single faults, the proposed approach further targets two fault modes that have received less emphasis in the existing literature: non-stationary single faults and mixed stationary faults. Dynamic adaptation of the sliding window size based on signal characteristics improves diagnostic accuracy across the machine's entire speed range while reducing latency. A range-based evaluation strategy is proposed to assess the dynamic diagnostic performance when multiple operating states coexist within the same sliding window. First, machine learning models trained on multi-scale experimental current data demonstrated robust fault diagnosis under various operating conditions. Second, a dynamic scaling factor for window size calculation was developed, incorporating signal metrics such as dominant frequency proportion, signal variation rate, and signal-to-noise ratio, enabling real-time adjustment of window sizes based on current information and short-term prediction. Third, offline optimization determines speed-specific parameter combinations, facilitating dynamic self-tuning of ASWIN parameters according to variable machine speed. Experimental results validated the proposed strategy, achieving 97.70% accuracy with a diagnosis time of approximately 60 ms, corresponding to a range of 0.4-2.4 current fundamental periods across a speed range of 200-1200 rpm.