DSLNET: dual-path SE-LSTM network for robust safe injection pump fault diagnosis under strong noisy environment

Safe injection pump fault diagnosis presents considerable challenges due to severe background noise and fluctuating operating conditions. Traditional diagnostic methods demonstrate limited feature extraction capability and high sensitivity to noise, while deep learning approaches face analogous constraints. To overcome these constraints, this study introduces DSLNet, an innovative diagnostic architecture for failure diagnosis. The proposed architecture first employs dual-path parallel convolutional pathways featuring complementary kernel sizes (large and small) to facilitate multi-scale feature extraction. Subsequently, multiplicative feature fusion is utilized to enhance complementary information and minimize redundancy. Squeeze-and-excitation modules with channel attention are then incorporated to automatically emphasize critical features and suppress noise. Furthermore, bidirectional LSTM layers are integrated to capture temporal dependencies. Validation on safety injection pump fault datasets confirms that DSLNet delivers exceptional robustness under high-noise operating conditions.