【摘要】 带式输送机托辊在长期运行中易受生锈、磨损等故障影响，其早期故障声学信号微弱且易被工业环境强噪声淹没，传统诊断方法存在特征提取不足与鲁棒性差的问题。文章提出一种基于CNN-Swin Transformer双分支特征融合网络(CSF)的故障诊断方法。通过融合变分模态分解与快速傅里叶变换构建时频域特征矩阵，结合CNN的局部特征提取优势与Swin Transformer的全局注意力机制，设计SE-CGA注意力机制实现深度特征提取。实验表明，该方法在真实工业数据集上达到98.11%的测试准确率，较单一CNN模型性能提升超过9%。在叠加-15 dB极端噪声时仍保持65.59%的识别精度验证了其在强噪声场景下的诊断鲁棒性与工程应用价值。更多还原

【Abstract】 Idlers of belt conveyors are susceptible to faults such as rust and wear during long-term operation. Their early-stage fault acoustic signals are weak and easily masked by strong industrial environmental noise. Traditional diagnostic methods suffer from insufficient feature extraction and poor robustness. We proposed a fault diagnosis method based on a CNN-Swin Transformer dual-branch feature fusion network(CSF). By integrating Variational Mode Decomposition(VMD) and Fast Fourier Transform(FFT), a time-frequency domain feature matrix was constructed. Combining the local feature extraction advantages of CNN with the global attention mechanism of Swin Transformer, an SE-CGA attention mechanism was designed to achieve deep feature extraction. Experiments demonstrated that the proposed method achieved a test accuracy of 98.11% on a real-world industrial dataset, representing an improvement of over 9% compared to a single CNN model. It maintains a recognition accuracy of 65.59% even under extreme noise conditions with a-15 dB signal-to-noise ratio, validating its diagnostic robustness in strong noise scenarios and its engineering application value.更多还原