Acoustic–vibration fusion bearing fault diagnosis via a multi-scale Swin–CNN hybrid architecture

Abstract

To address the problems that weak impulsive features in acoustic–vibration signals are easily masked under strong noise, fault severities within the same fault category are difficult to distinguish, and existing fusion models show insufficient coordination between local details and global semantics, this paper proposes an acoustic–vibration fusion method for bearing fault diagnosis based on a multi-scale Swin–CNN hybrid architecture. The proposed method first employs a Bayesian optimization-based tunable Q-factor wavelet transform (BO-TQWT) to enhance fault-sensitive subbands under low signal-to-noise ratio conditions, and then converts acoustic and vibration signals into two-dimensional time–frequency maps. Subsequently, a Swin–CNN hybrid network is constructed, in which CNNSwinBridge performs a two-stage, single-pass cross-guided recalibration between CNN-derived local features and Swin-derived contextual features. Specifically, Swin features provide channel-wise semantic guidance for CNN features, whereas the recalibrated CNN features provide spatial texture guidance for Swin features. The module provides lightweight cross-architecture coordination without recurrent or iterative feedback. Experimental results on the BJTU-RAO and University of Ottawa bearing datasets show that the proposed method achieves average accuracies of 99.51% and 99.96%, respectively, under clean operating conditions. Further analysis indicates that BO-TQWT provides relatively limited performance gains under clean conditions, whereas it can more effectively enhance fault-sensitive features under strong noise, thereby improving the fine-grained discrimination of different severity levels within the same fault category.