Fault diagnosis of rotating machinery based on combination of Wasserstein generative adversarial networks and long short term memory fully convolutional network

Abstract

The traditional fault diagnosis methods of rotating machinery based on deep learning have made some achievements. However, the fault samples are generally difficult to collect, which will lead to data imbalance problem. In such cases, it will degrade their diagnosis performances. Therefore, a data rebalancing solution based on Wasserstein Generative Adversarial Network (WGAN) was proposed to improve data quality, and combined with long and short-term memory full convolutional network (LSTM-FCN) to achieve high-dimensional vibration signal fault diagnosis. The method first utilizes WGAN to expand the imbalanced fault samples. The unique one-dimensional convolution architecture in WGAN and the loss function based on Wasserstein distance ensure the stable feature learning ability of the model and the high quality of the generated data. Hence, it fully solves the problem of data imbalance. Subsequently, the balanced high-dimensional fault signals are fed into LSTM-FCN model. LSTM-FCN simultaneously combines the characteristics of LSTM and FCN in timing feature extraction, and thus extracts more robust and representative fault features through parallel feature extraction and fusion technology. Finally, SoftMax classifier is employed to identify different fault conditions. The proposed method is verified on the multiple unbalanced fault datasets of rolling bearings and gears. Compared with the traditional data imbalance resolution algorithms, the experimental results demonstrate that WGAN can effectively deal with the data imbalance problem, and the diagnosis accuracy of LSTM-FCN is maintained at about 99%. In comparison with the traditional fault diagnosis algorithms, the proposed method shows better diagnostic performance and generalization ability under complex working conditions with small samples and strong noise.