You can be more trustworthy: A feature fusion reinforcement network for credible anti-noise fault diagnosis

Significant research progress has been made in intelligent fault diagnosis algorithms. However, these methods face challenges such as noise interference and untrustworthy diagnostic results in industrial practice, which limit their performance in practical applications. This paper proposes a new feature fusion and reinforcement network combined with swin-transformer (Swin-FFRN) for noise-resistant and trustworthy diagnosis, which combines a global feature extraction network and a staged convolutional fusion operation for fine-grained fault feature extraction and noise suppression in 2D time–frequency maps. The Swin-FFRN is used to analyze 2D time–frequency map data of different mechanical faults in a low signal-to-noise ratio environment by introducing a channel attention mechanism and a spatial attention mechanism to strengthen the critical fault features that are strongly correlated with the classification of the faults so that the model focuses on the crucial features. Moreover, the noise immunity performance is evaluated using the latest methods on two different datasets, and intuitive visual interpretability is provided to show model credibility. The results show that the noise-resistant diagnostic accuracy of the proposed method is improved by 5.43% on average with respect to the SOTA method. By enhancing the key input features, the proposed method can give diagnostic results with a reasonable decision basis.