Denoising Mixed Attention Variational Auto-encoder for Axial Piston Pump Fault Diagnosis

As an energy supply component of hydraulic system, the fault diagnosis of axial piston pump is of great significance. However, most existing methods rely on expert knowledge for feature extraction, and the robustness to noise is poor. To tackle the problem that complex working conditions bring noise interference to the collected diagnostic signals of axial piston pump, an end-to-end denoising mixed attention variational auto-encoder method is proposed to directly extract the fault characteristics submerged in the noise, to realize the fault diagnosis of axial piston pump under noisy environment. The proposed method employs convolution variational auto-encoder to extract fault features from multivariate signals including pressure and vibration. By introducing the mixed attention mechanism, hidden layer features of the encoder are weighted and fused, enhancing the fault features while weakening the noise. The adaptive soft-threshold denoising method is further applied to reducing the noise interference in extracted features, realizing the fault diagnosis of axial piston pump under strong noise. The effectiveness of the proposed method is verified by the fault implantation experiment and noise robustness experiment of an axial piston pump, and the results show 99.32% diagnosis accuracy under 5 dB noise and 69.72% under −5 dB noise, which outperforms commonly used diagnosis methods. © 2024 Chinese Mechanical Engineering Society. All rights reserved.