Fault State Identification Method with Noise Robustness of Dry Gas Seals Based on Sample-Augmented MA1D-ResNet

Dry gas seals are widely used in the petrochemical industry for shaft end sealing of compressors, pumps and other equipment involving flammable, explosive, toxic and harmful media. To address the challenge of accurately identifying the fault states of dry gas seals under strong noise interference, this paper proposes a Multi-scale Attention 1D Residual Network (MA1D-ResNet) model based on sample augmentation. First, a dry gas seal acoustic emission (AE) test rig was built to collect non-stationary AE signals. The training dataset was expanded to five times its original size through data segmentation and Gaussian noise injection, significantly enhancing the model's generalization capability in the data-input domain and training process. Then, the proposed model incorporates a Multi-scale Dual Attention Module (MDAM) into the ResNet18 architecture: it employs 1D convolutions to process temporal signals directly, avoiding feature loss, and integrates MDAM after the first convolutional layer and the Stage1 layer to strengthen fault feature extraction. Finally, experimental results demonstrate that the proposed model achieves an average accuracy of 99.8571% in classifying seven fault states (significantly outperforming five comparative models including CNN, ResNet, and ResNet-CBAM), with 100% recognition rate for five of the fault categories. The proposed model exhibits outstanding noise robustness, maintaining an accuracy of 92.43% under strong noise conditions of -6 dB. This study provides a highly robust solution for the intelligent fault diagnosis of dry gas seals in complex noise environments.