Research on Diaphragm Pump Fault Diagnosis Method Based on Res-DCB-Net

Abstract

Nonstationary pressure pulsation signals of diaphragm pumps contain strong background noise and coupled characteristics. This makes it challenging to extract incipient fault features and to decouple faults with similar physical mechanisms. To address these limitations, this paper proposes a spatiotemporal fault diagnosis model named Res-DCB-Net. A dynamic principal component analysis (DPCA) algorithm is used to reconstruct one-dimensional nonstationary signals into two-dimensional time–frequency spectra with a high signal-to-noise ratio, eliminating fluid background redundancy. A deep residual network (ResNet) incorporating a convolutional block attention module (CBAM) is constructed to adaptively focus on core transient impact textures in the spatial dimension while mitigating gradient degradation. A bidirectional long short-term memory (BiLSTM) network is integrated to capture the long-range temporal evolution of feature sequences over the reciprocating cycle. Evaluated on a dataset from an industrial diaphragm pump fault simulation platform, the results show that Res-DCB-Net achieves a recognition accuracy of 99.26%. Compared with standard baselines (SVM, 1D-CNN, and ResNet-18), the accuracy is higher by 4.54%, 14.64%, and 5.55%. Compared with recent architectures such as Vision Transformer (ViT) and MobileNetV3, the proposed model maintains higher precision and reduces interclass confusion, thereby breaking through the classification limitations of highly similar incipient faults.