Abstract: In response to the issues of high signal complexity and insufficient spatial feature extraction capability in hydraulic pump fault diagnosis,a fault diagnosis method combining CloFormer and Bi-LSTM was proposed. This method embedded the Bi-LSTM neural network into the CloFormer network to enhance the model′s learning capability of fault features and improve the representation performance of global long-distance information and local temporal information. Additionally,a multi-layer feature fusion structure was designed to learn and integrate multi-scale features,promoting the flow and transfer of feature information. Finally,this proposed method was applied to the task of hydraulic pump fault diagnosis to verify its effectiveness. The results demonstrate that the proposed method achieves accurate diagnosis of 5 categories of hydraulic pump states,with accuracy and F₁ scores reaching 98.74 % and 98.85 % respectively,outperforming other 5 comparative methods.

Key words: hydraulic pump, fault diagnosis, CloFormer, bidirectional long short-term memory network, global-local information