A Bearing Fault Diagnosis Model Based on SCSA-CNN-LSTM with Linear Deformable Convolution

To overcome the constraints of conventional rolling bearing fault diagnosis approaches, which often extract insufficient features and struggle to deeply mine fault characteristics under complex operating conditions, this paper proposes a CNN-LSTM model based on linear deformable convolution (LDConv) and spatial and channel synergistic attention (SCSA) mechanism. The proposed model uses CNN as the backbone network, efficiently extracting features by incorporating LDConv and combining the SCSA module to uncover hidden channel and spatial information. Then, LSTM is utilized to leverage its sequence learning capability to extract the temporal features of the signal, ultimately achieving the identification and classification of bearing fault states. Experiments conducted on the Case Western Reserve University dataset under various noise conditions demonstrate that the proposed model effectively extracts fault features from the raw signals and exhibits strong robustness. The model achieves an accuracy of 95.86% in noisy environments, significantly outperforming other advanced fault diagnosis models. Furthermore, ablation studies confirm the effectiveness of the proposed modules.