Research On Transfer Fault Diagnosis Method Of Rolling Bearings Based On Deep Domain Adversarial Neural Network

In recent years,deep neural networks have been successfully applied to bearing fault diagnosis due to their end-to-end modeling and adaptive feature extraction.Deep learning techniques are more dependent on the amount of training data.However,in actual engineering applications,due to various factors,there are often phenomena such as insufficient fault data and lack of effective labeled data.Such phenomenon tends to cause model deviations,which reduces the accuracy and stability of the diagnosis results.As a result,the application effect of deep learning techniques in bearing fault diagnosis is seriously restricted.To solve the above problems,this paper introduces deep transfer learning techniques into bearing fault diagnosis problems,improves the fault diagnosis effect of target bearings by using auxiliary data under different working conditions.Specifically,this paper adopts deep domain adversarial neural network as basic model,takes structured information among multiple bearing fault conditions as the breakthrough point.Aiming at two situations of whether the data of target domain is labeled,constructed semi-supervised and unsupervised deep domain adversarial neural network models,respectively.The proposed methods can realize the effective transfer of fault information between different working conditions,which solves the accuracy of deep learning model decreases and numerical instability problems caused by insufficient target data.Meanwhile,the accuracy and robustness of fault diagnosis model have also been greatly improved.The main research contents include:（1）Bearing vibration signals have the characteristics of time series,noise,periodic fluctuations,etc.Therefore,it is necessary to study the applicability and modeling method of deep adversarial training mechanism to vibration signal data before building a deep domain adversarial neural network.This paper adopts a typical deep adversarial training model,namely generative adversarial network（GAN）,combined with stacked denoising auto-encoder to construct an imbalanced fault diagnosis model based on deep adversarial training mechanism.The result of simulations experiment on CWRU bearing fault data set published by Case Western Reserve University in the United States has shown that: the GAN based on the adversarial training mode can effectively extract the data distribution characteristics and improve the quality of virtual sample synthesis.However,there are still deficiencies in the diagnosis problems with large distribution differences and multiple fault types.（2）Focused on the problem that only a small amount of labeled data in target domain and insufficient stability problems,this paper proposes a semi-supervised structured domain adversarial neural network（SDANN）for rolling bearing transfer fault diagnosis.First,construct a discriminative regularizer via maximum correlation entropy constraint to improve the discriminative ability of multiple fault types.Second,introduce symmetry constraint of the structured information matrix to capture the intrinsic similarity information among multiple fault types.Finally,realize semi-supervised transfer fault diagnosis via the built SDANN.Comparative experiments are conducted on two widely-used bearing datasets CWRU and XJTU-SY.The results show that the proposed method has good diagnosis performance on insufficient monitoring data and outperforms eleven state-of-the-art methods with and without transfer techniques in terms of diagnostic accuracy and numerical stability.（3）Based on the above work,this paper further considers the situation that the target working condition data has no label information at all,proposes an unsupervised structured domain adversarial neural network for transfer fault diagnosis.This method firstly introduces pseudo-label techniques to DANN so as to obtain the pseudo-labels of unlabeled data in target domain.Secondly,add structured relationship matrix among multiple health conditions in the source domain and the target domain,respectively.Thirdly,construct the structured information symmetric constraint of the two domains to make the structured information of the two domains as close as possible to improve the quality of pseudo-labels.Finally,in order to achieve rapid convergence of pseudo-label and domain adversarial training,this paper also introduces the MMD distance.This method iteratively optimizes the structured domain adversarial training process and the pseudo-label generation process,and finally obtains an end-to-end unsupervised transfer fault diagnosis model.The experimental results on CWRU dataset show that the proposed method can obtain effective and stable transfer fault diagnosis results.In summary,this paper is oriented to the characteristics of actual engineering.The fault diagnosis model performs better in accuracy and robustness by improving the deep transfer learning model.It provides a new solution for fault diagnosis when there are few or no fault data labels,which has significant academic research and engineering application value.