Methodologies For Fault Diagnosis Of Rotary Machine Considering Domain Differences

Fault diagnosis technology,as a key technology to guarantee the reliability,safety and maintainability of equipment,has received wide attention at home and abroad.The data-driven fault diagnosis method can realize real-time and intelligent fault diagnosis of equipment through mining and analysis of big data,which has become a hot spot of current research.In recent years,many achievements have been made in the field of fault diagnosis based on data-driven methods,but there are still the following problems that need to be solved:(1)When a fault occurs in the equipment,the vibration signal collected by the sensor to characterize the operation status of the rotating machinery and equipment is nonlinear,non-smooth and weakly periodic,and it is difficult to accurately extract the key features that can reflect the health status of the equipment.(2)Although the fault diagnosis method based on deep learning can automatically extract fault features end-to-end,it requires a large number of training samples,and the training model contains a large number of weight parameters and biases,which leads to the application effect of the method being greatly affected by the number of samples and a long training time.(3)The existing fault diagnosis methods assume that samples follow a similar distribution.However,actual engineering machinery and equipment work in normal state most of time and their fault samples are rare.In addition,there are great differences in equipment models,operating conditions and fault degrees,it is difficult for the fault samples to obey same distribution.In summary,the research on cross-domain fault diagnosis of rotating machinery is of great significance to practical engineering.To solve the above problems,bearings,gears and electric spindles are used as application objects,and fault diagnosis method based on the data-driven is as a theoretical basis to carry out research.(1)The combined EEMD and MPE are used as feature extractors,and the combined GA and BP are used as pattern classifiers to propose a double-optimized machine learning method.By reasonably selecting the parameters and models of the method,the extraction of fault features from fault signals with nonlinear,non-smooth,and weakly periodic characteristics is achieved,and the accuracy of fault identification is improved.First,the key feature nodes characterizing the component health state are extracted from several different scales by combining EEMD and MPE methods to build the feature vector set.Then,the BP method is used for the classification and identification of features.Among them,the initial parameters of the back propagation algorithm are selected and optimized by a genetic algorithm.It is demonstrated that the method can achieve accurate assessment in different cases.(2)To address the problems of deep learning-based fault diagnosis method have many internal parameters,high training time cost and degradation of diagnosis accuracy under variable working conditions,we propose a fault diagnosis method of convolutional neural network with few parameters and domain adaptation capability.The influence of the number of training parameters required by the deep learning model on the calculation time and the interference of the difference between the source domain and the target domain on the model stability are considered comprehensively.The structure of the convolutional neural network is optimized based on five lightweight strategy,and the output of the batch normalization layer is optimized using an adaptive batch normalization method,so as to achieve highly accurate and low-cost crossdomain fault diagnosis.The robustness and adaptivity of the proposed method are verified using bearing fault data,gear fault data and electric spindle fault data.(3)The research on the cross-domain fault diagnosis method of non-identical distribution of single source domain is carried out.In view of the problem that the assumption of same distribution is difficult to be established due to different load conditions of mechanical equipment in practical engineering,a deep convolutional neural network with multi-layers is used as the basis for cross-domain fault diagnosis.By learning the common knowledge contained in the fault samples in the source domain,the generalization performance of the proposed method for the target domain from different load conditions is improved,and the domain adaptive ability of the crossdomain fault diagnosis method is also improved Combining the optimized convolutional neural network architecture,Ada BN and MKMMD to adjust statistical information of all batch normalization and distance of the high-dimensional feature vectors from the source domain and the target domain,the domain self-adaptation of the proposed method is realized.Finally,the fault data of two kinds of faulty bearings are used to verify the applicability of the proposed method.(4)To solve the problem that it is difficult to obtain a large amount of singlecondition marked fault data from the same mechanical equipment,a multi-source domain cross-domain fault diagnosis method based on MCD is proposed.By adversarial learning,the diagnostic features that have the ability of pattern discrimination and are not interfered by domain change are extracted from the fault data of the same equipment with multiple working conditions or the same working principle but different equipment sources,and then applied to the target domain.Since,the domains of data have obviously different characteristics,the feature extractor and the convolutional neural network with double classifiers are used for adversarial training,the absolute value of the difference between the probability output of the source domain and the target domain samples in the two classifiers is minimized as the project.Then,the feature extractor can extract domain invariant features from multiple source domains and target domains,and the classifier can recognize the target task pattern accurately.Loss functions optimized for conditional entropy and KL scatter are designed to force the decision boundary of the classifier away from data-intensive regions,thus improving the accuracy of fault diagnosis.The advantages of the method are verified by performing applications on bearings and gears.