Research In Mechanical Fault Diagnosis Methods Based On Generative Adversarial Network

Mechanical equipment is important to monitor the status and diagnose the fault of the equipment so as to ensure its safe operation.With the development of computers,sensors and communication technologies,the condition monitoring of electromechanical products has entered the era of "big data" and brings out new challenges such as huge data volume,many data types and high speed.In order to improve the accuracy and efficiency of fault diagnosis and prediction,the new theories and new methods of big data for machine fault diagnosis are the hot spots of research currently.The new generation of artificial intelligence technology based on deep learning is an effective method to realize intelligent fault diagnosis through feature of data mining deeply and knowledge independent-study.It is of great theoretical significance and practical value to use new theories and new methods for the on-line motoring and diagnosis.In reality,it is difficult to acquire enough fault samples because normal condition data are more prevalent than faulty condition data in real manufacturing environments,which limits the accuracy and stability of the diagnosis.Generative Adversarial Network(GAN)is a branch of deep learning and may be the most valuable method in dealing with the imbalanced datasets.Aiming at improving accuracy rate and reliability of diagnosis and taking the key parts of mechanical equipment as the detection object,the fault generation technique and diagnosis method of the mechanical equipment based on GAN are studied.A new generation method is proposed to solve the small samples and imbalanced datasets in machine fault diagnosis.In this study,based on the characteristics of Conditional Generative Adversarial Network(CGAN)and Deep Convolutional Generative Adversarial Network(DCGAN),the combined model Conditional Deep Convolutional Generative Adversarial Networks(C-DCGAN),in which the generator and discriminator are both Convolutional Neural Networks(CNN),is adopted.Generator captures the true distribution of the original vibration data and generates new samples with similar distribution to original data to enrich datasets to balance the fault data.Discriminator discriminates input samples from true samples.The generated new samples together with the original samples are imported to the discriminator to improve the generalization ability of the fault classifier.During the GAN training process,the discriminator CNN and the generator CNN are alternately optimized by the anti-learning mechanism to improve the quality of the generated samples.Then another CNN model is separately trained for fault recognition.Case Western Reserve University(CWRU)bearing dataset and the measurement dataset in the laboratory were used to validate the proposed method.In order to solve the training instability,a two time-scale update rule(TTUR)for GANs is introduced.We propose using TTUR specifically to compensate for the problem of slow learning in a regularized discriminator,making it possible to use fewer discriminator steps per generator step,which makes C-DCGAN training more stable.Jensen-Shannon divergence(JSD)which captures the similarity of generated data to real ones is used to evaluate the performance of TTUR on C-DCGAN.We have compared C-DCGAN trained with TTUR to conventional C-DCGAN training with a one time-scale update rule on CWRU and planetary gear box datasets.TTUR outperforms conventional C-DCGAN training consistently in experiments.The vibration signal of mechanical equipment has the complexity of frequency component and time-varying.To solve this problem,self-attention mechanism is introduced to C-DCGAN model,C-DCGAN fault diagnosis method based on self-attention mechanism(SAC-DCGAN)is proposed.The feature attention distribution was calculated for all input vibration signals,and the weight between the output feature and the input feature was calculated,that is,the attention degree to the feature vectors of input vibration signal.Then the weighted average of feature information is calculated according to the attention distribution.The vibration characteristics output by the convolutional layer at different times are dynamically weighted,so that the fault model can adaptively focus "attention" at different times when there is a significant difference in time characteristics,and solve the sample difference caused by the time-varying of vibration signals.Finally,the proposed method is validated on the planetary gearbox dataset and we adopt Dynamic Time Warping(DTW)to evaluate the quality of the generated samples.The visual experiment result indicates that the proposed SAC-DCGAN performs better than C-DCGAN,which can accurately diagnose various states planetary gearbox.Finally,the fault diagnosis model SAC-DCGAN is evaluated and validated as a whole.Three evaluation metrics of the model are proposed for fault vibration signal: Jensen-Shannon divergence(JSD),Kernel Maximum Mean Discrepancy(MMD)and the 1-Nearest Neighbor classifier(1-NN),which were calculated to evaluate the generator’s ability to simulate the distribution of training data.Meanwhile,we adopted an evaluation method based on samples.SAC-DCGAN was applied in the experiments to test the three abilities of evaluation metrics:three abilities to distinguish generated samples from real samples,test mode collapsing and detect overfitting.The proposed scheme can effectively improve diagnosis accuracy and generalization ability in dealing with issues including few-shot learning and imbalance of the data set,providing new solutions and ideas for fault diagnosis of machinery in “big data” and guiding fault diagnosis to develop on scope or on depth in the practical project application.