Fault diagnosis method of mine hoist main bearing with small sample based on VAE-WGAN

Abstract

As a key component of the hoist, the main bearing may deteriorate in performance and cause faults during long-term high-speed and heavy-duty service. Therefore, conducting fault diagnosis of the main bearing of the hoist is of great significance for ensuring the safe and efficient operation of the mine hoist. However, the proportion of normal service and fault status in the monitoring data of the operation status of mine hoists is severely imbalanced, showing characteristics such as a large number of normal samples, a small number of fault samples, and insufficient label samples, resulting in unsatisfactory training results and low diagnostic accuracy of the main bearing fault model of mine hoists. In response to the problem of low accuracy in fault diagnosis of mine hoist main bearings under small samples, a VAE-WGAN based mine hoist main bearing sample augmentation model was constructed by fusing variational autoencoder and Wasserstein to generate adversarial networks. Furthermore, a fault diagnosis method based on CBAM-MobileNetV2 is proposed to achieve fault diagnosis of mine hoist main bearings under small sample data. At the algorithmic level, the Wasserstein distance metric is introduced to solve the problem of vanishing training gradients in generative adversarial networks. At the data level, VAE-WGAN was tested using the Case Western Reserve University dataset, and hyperparameters were optimized by evaluating the generation ability of VAE-WGAN through quantitative indicators. VAE-WGAN was trained using the bearing dataset of the mine hoist simulation test bench to achieve the expansion and expansion of the small sample dataset. In order to improve the feature extraction ability and fault diagnosis accuracy of fault diagnosis models, based on the lightweight convolutional neural network MobileNetV2, the convolutional block attention mechanism CBAM is integrated into the deep feature mapping of MobileNetV2, and an attention mechanism convolutional classification network CBAM-MobileNetV2 is constructed. By integrating cross channel and spatial information, more attention is paid to fault features. Finally, a comparative analysis was conducted with traditional generation models such as WGAN-GP, DCGAN, VAE, and WGAN. The fault diagnosis accuracy of VAE-WGAN+CBAM-MobileNetV2 on four small sample ratio datasets was higher than the other four methods, proving that the proposed sample augmentation and fault diagnosis methods have higher diagnostic accuracy on different small sample ratio fault datasets and can meet the fault diagnosis requirements under small samples.