Research on Cavitation Fault Diagnosis of Axial Piston Pumps Based on Rough Set Attribute Weighted Convolutional Neural Networks

TL;DR

This paper introduces a new method for diagnosing cavitation faults in axial piston pumps using a combination of rough set theory and convolutional neural networks, achieving high accuracy and interpretability.

Contribution

The novel RSAW-CNN method integrates rough set attribute weighting with CNNs for improved cavitation fault diagnosis in piston pumps.

Findings

The RSAW-CNN method achieved an average diagnostic accuracy of over 99.2%.

It outperformed other neural network models by 5.40% to 15.87% in accuracy.

The method provides interpretable decision support through extracted decision rules.

Abstract

Cavitation phenomenon in piston pumps not only causes vibration and noise but also leads to component damage. Conventional diagnostic methods suffer from low accuracy, while deep learning approaches lack interpretability. To address these limitations, this paper proposes an intelligent fault diagnosis method based on the rough set Attribute Weighted Convolutional Neural Network (RSAW-CNN). First, based on cavitation mechanisms and the mathematical model, the computational fluid dynamics model of the piston pump is established to simulate the failure condition. Subsequently, employing rough set theory, an original fault decision table is constructed, discretized, and subjected to attribute reduction. A weight matrix is generated according to the importance of each data channel in the classification decision and embedded into the input layer of the Convolutional Neural Network (CNN) to enhance the influence of key features. Decision rules are also extracted to provide interpretable decision support for fault diagnosis. Experimental results demonstrate that the proposed RSAW-CNN method achieves an average diagnostic accuracy of over 99.2%. Compared to the backpropagation neural network, residual neural network, CNN, and the CNN with squeeze-and-excitation networks, its average accuracy has improved by 15.87%, 10.83%, 7.48%, and 5.40%. The proposed method not only exhibits high diagnostic accuracy but also offers strong interpretability and reliability.

Figures