To address the issues of poor noise resistance and the lack of mechanistic analysis in the classification and diagnosis of fault vibration signals collected by sensors using existing deep learning models, this paper proposes a Transformer-based fault diagnosis model incorporating squared convolution under strong noise conditions, namely CS-Transformer. This model enhances the local feature representation of fault vibration signals through wide convolution kernels and squaring operations, improves the robustness of global features by leveraging global average pooling, and employs a single-layer Transformer encoder to uncover the correlations among global features, thereby further focusing on key fault features. Fault diagnosis experiments were conducted based on the CWRU and Paderborn bearing datasets. When the signal-to-noise ratio is -6 dB, the noise resistance of the model exceeds 91%, significantly outperforming other comparable models. This validates the superior classification performance and generalization ability of this model for bearing faults of varying degrees under strong noise conditions. Moreover, the analysis of the visualized envelope spectrum further confirms that this model can effectively enhance the target fault frequency and suppress the noise.
23 March 2026 in Eksploatacja i Niezawodność – Maintenance and Reliability
In response to the challenges of fault diagnosis for industrial equipment in high-noise environments, this study proposes an improved graph neural network model that combines envelope activation and cross-correlation function. This model first enhances single-channel feature extraction and anti-noise ability through an improved independent parallel one-dimensional convolution structure and envelope activation operation, achieving single-channel noise suppression and feature extraction; then, it uses the time-domain cross-correlation function to construct a topological graph structure between channels, highlighting the essential correlations of multi-channel signals, and realizes multi-channel feature fusion through graph convolution networks. Experimental results on the public bearing dataset show that proposed model has high classification accuracy, superior anti-noise performance, and good generalization ability in high-noise environments.