An interpretable multi-sensor bearing nonlinear composite fault diagnosis method for wind power generation systems

Driven by global clean energy strategies, wind power develops rapidly. Bearings, core wind turbine transmission parts, govern system reliability and safety. Conventional diagnosis suffers three key practical limitations: single-sensor signals cannot fully characterize nonlinear composite faults; mainstream deep learning models act as opaque black boxes without clear diagnostic interpretability; highly coupled composite fault features cannot be separately extracted by existing algorithms. To address these challenges, this paper proposes an interpretable multi-sensor bearing nonlinear composite fault diagnosis method for wind power systems. Firstly, a multi-sensor dynamic frequency guided synchronous compressed wavelet transform is designed to precisely extract and unify multi-sensor non-stationary signal features via dynamic frequency matching, adaptive wavelet basis selection, and scale parameter optimization. Secondly, a dynamic calibration and feature enhancement network is constructed, including a dynamic dual-branch calibration fusion module for adaptive feature weighting and decoupling, and a wavelet attention feature enhancement network for sensitive feature enhancement and interpretability improvement. Finally, a Mahalanobis distance aware Krylov Transformer network is developed, integrating Mahalanobis distance to enhance early subtle fault sensitivity and an efficient global enhanced Krylov Transformer for deep feature modeling. Experiments across the three datasets yield average diagnostic accuracies of 98.80, 99.24, and 99.93%, respectively. Even under −4 dB noise interference, the proposed model retains an average accuracy above 90%. Component decoupling verification reveals that 96.0% of composite fault samples can be simultaneously identified via two independent fault channels, with the Pearson correlation coefficient between channel outputs as low as 0.18. Moreover, controlled sub-band masking tests show that masking the HH sub-band containing fault impulse information reduces the overall diagnostic accuracy from 98.45 to 76.89%, corresponding to a 21.56 percentage point drop, this sufficiently proves that the model’s inference relies heavily on high-frequency time–frequency features corresponding to fault impulses.