Stage-aware multi-sensor time-frequency measurement and lightweight deep learning for mechanical fault diagnosis of OLTC

Abstract Timely detection and diagnosis of abnormal conditions in on-load tap-changer (OLTC) are critical for maintaining the stable operation of power grids, yet in practical monitoring scenarios the vibration signals are highly non-stationary and often contaminated by strong noise and coupled mechanical responses, making conventional approaches prone to stage mixing, incomplete utilization of spatially distributed measurements, and blurred time–frequency representations. These issues, frequently observed in real OLTC datasets, directly reduce feature separability and lead to unstable diagnostic performance, especially under multi-fault or small-sample conditions. In response to these challenges, a stage-aware multi-sensor collaborative diagnostic framework is developed, where spatially distributed vibration measurements are leveraged to capture differentiated responses from the drive mechanism and main body, allowing complementary fault perception. To better align signal processing with the physical operation of OLTC, a stage segmentation strategy is introduced by integrating multi-scale Teager energy operators with motor current signals, enabling the isolation of fault-relevant segments while suppressing interference from irrelevant operating stages. Building on this, a mechanically informed MPEW-CEEMDAN method is designed to adaptively regulate the decomposition process, improving robustness against mode mixing and noise, while the reassignment method further sharpens the time–frequency representation by concentrating dispersed energy into physically meaningful regions. The resulting time–frequency matrices extracted from the three vibration measurement points are organized into a three-channel tensor, preserving both spatial diversity and time–frequency characteristics for subsequent learning. To ensure applicability in resource-constrained environments, a lightweight MobileNetV4–ResNet-18 hybrid network is constructed, achieving a balance between computational efficiency and deep feature representation. Experimental results demonstrate that the proposed framework significantly enhances the interpretability of time–frequency measurements and the separability of OLTC operating states under complex mechanical backgrounds, ultimately achieving highly reliable fault diagnosis.