Study on compound fault diagnosis of rolling bearings founded on ACTSR-SVMD and a two-branch parallel network

Abstract

To resolve the challenge of isolating complex fault features in rolling bearings amidst multi-source interference, which leads to diminished diagnostic accuracy, this study proposes an adaptive cascade three-stable stochastic resonance (ACTSR) and successive variational modal decomposition (SVMD) signal processing framework. Furthermore, we develop a hybrid diagnostic methodology that integrates a two-branch parallel network (1DCNN + Informer) with a multi-head attention-based feature fusion mechanism. First, we first optimise the parameters of ACTSR-SVMD, then decompose and denoise the bearing vibration signals to extract multiple Intrinsic Mode Functions (IMFs). Effective IMF components are selected based on correlation coefficients for signal reconstruction. Second, local time-domain features are extracted using 1DCNN, while Informer captures global temporal dependencies. Multi-scale feature fusion and fault identification are achieved by combining multi-head attention and dynamic weighting. On a laboratory fault dataset, the proposed model achieves a diagnostic accuracy of 99.6% with a misclassification rate of approximately 0.31%, outperforming five comparative models, including CNN-Attention. When tested on the public XJTU dataset, the model achieves leading diagnostic accuracy across all sample types, with a misclassification rate of less than 2% for different diagnostic tasks. The results demonstrate that the proposed method significantly enhances the accuracy and reliability of rolling bearing fault diagnosis.