Fault diagnosis of vane pumps based on optimized noise reduction and attention mechanism ResNet fusion

Abstract

Addressing issues such as susceptibility to noise interference and limited diagnostic accuracy in complex operating conditions for vane pump fault characteristics, this paper proposes a fault diagnosis method for vane pumps based on optimized noise reduction and attention mechanism ResNet fusion. Firstly, an experimental platform was established to collect raw vibration signals from vane pumps under normal operation and six typical wear faults. The SABO-VMD method was employed for signal denoising and reconstruction. Subsequently, the reconstructed signals undergo CWT time-frequency analysis to construct a fault feature sample set. Finally, two enhanced ResNet models—SE-ResNet and CBAM-ResNet—are developed for fault diagnosis. Results demonstrate that the SABO-VMD algorithm achieves an 11.34% and 13.23% improvement in signal-to-noise ratio compared to traditional EEMD and LMD methods, respectively, alongside a 19.56% and 23.04% reduction in root mean square error. It also outperforms PSO-VMD and GA-VMD. Under noise-free conditions, both SE-ResNet and CBAM-ResNet achieve 100% accuracy. Under noisy conditions, SE-ResNet (97.07%) and CBAM-ResNet (97.57%) significantly outperformed the baseline ResNet (90.60%), 2D-CNN (96.29%), and VIT (94.57%) in diagnostic accuracy. The two proposed enhanced ResNet models effectively extract blade pump fault features amidst strong noise backgrounds, markedly improving the diagnostic model’s interference resistance and recognition accuracy.