Conditional Feature Generative Adversarial Network for Fault Diagnosis of Axial Piston Pump

Abstract

A conditional feature generative adversarial network (CFGAN) for small-sample data augmentation was proposed in this article to address the issue of the scarcity of fault samples in axial piston pump fault diagnosis. With 1-D data as input, a feature extractor was constructed in this method to extract features, bringing about an original deep feature dataset. Subsequently, this dataset was employed to train the generation model for enhancing the training efficiency. Additionally, the Wasserstein distance was introduced to improve the stability of the model training. Conditional labels were added to the inputs of the generator and discriminator to guide the model’s training and generate high-quality feature datasets. Afterward, the effectiveness of the proposed data generation method was verified through the data obtained from the Paderborn Bearing Dataset and an axial piston pump fault simulation testbed. The fault diagnosis accuracy of the augmented dataset was significantly improved compared to the nonaugmented dataset. Furthermore, a multiscale convolutional neural network (MSCNN) was established and utilized different-sized convolutional kernels to extract and fuse multiscale information from the samples. This approach enabled the capture of more comprehensive fault state features, contributing to enhanced diagnostic accuracy. The experimental data of the axial piston pump validate that the diagnostic accuracy with fused multisensor signals was higher compared to a single sensor.