Generative Adversarial Network Based Methods For Rolling Bearing Fault Diagnosis

Nowadays,deep learning based fault diagnosis methods have been widely applied.For convolutional neural network based methods in fault classification,the more labeled data,and the better performance.However,a large amount of data is hard to get for labeling must be tagged by hand,and to get labeled data in a huge amount is costly in practice.And in this background out comes two problems: how to solve fault diagnosis with limited labeled data in a supervised way,and how to solve fault diagnosis with limited labeled data under a semisupervised way.In this paper,we solve these two problems based on generative adversarial network and conduct some experiments to validate the effectiveness of our proposed methods.A pre-processing method is also studied to guarantee the good performance.To validate that generative adversarial network can be suitable for fault diagnosis,a fault-data-based method is proposed.A signal-to-image method is used as data preprocessing.With 1D vibration signals converted into 2D images,more presentative features can be later extracted by convolutional neural network.The result on Case Western Reserve University standard dataset achieves 99.996%,validating the effectiveness of the proposed method.To solve fault diagnosis with limited labeled data in a supervised way,the ability of generation of generative adversarial network is adopted to imitate and generate data with similar distributions of the real input labeled data.Through this way,the dataset for classification can be enlarged and the final accuracy can be promoted.Moreover,to mitigate the instability brought by the decrease of labeled data between generator and discriminator,a threshold-control method is also proposed to adjust the relationship of generator and discriminator dynamically and automatically.The result on Case Western Reserve University standard dataset achieves 99.96%,validating the effectiveness of the proposed method.To solve fault diagnosis with limited labeled data under a semi-supervised way,unlabeled data with their information are also used to improve the classification performance.The input of discriminator consists of three parts: fake samples generated by generator,unlabeled real data,and labeled real data.The use of unlabeled data can be realized by the expanding of classification,which will be expanded from K to K+1.And the loss function consists of two parts: the supervised part for classification,and the unsupervised part to promote the classification with its information of unlabeled data.The result on Case Western Reserve University standard dataset achieves 100%,validating the effectiveness of the proposed method.Finally,we make a conclusion of our major contributions and list some future works which are worth studying.