Research On Several Kinds Of Fault Diagnosis Algorithms Based On Deep Learning

With the development of society,fault diagnosis has become an important link in modern industrial production process.The rapid development of information technology and artificial intelligence makes the fault diagnosis methods based on industrial data more and more competitive.However,the traditional data-driven fault diagnosis task still has many problems,such as the vulnerability to sample size,noise disturbance and other problems,which are important factors restricting fault feature extraction and fault diagnosis.Aiming at the problems of noise interference,data sample labeling,zero sample data and so on,this paper proposes several fault diagnosis methods based on deep learning.The main contents of this paper are as follows:First of all,the traditional data-driven fault diagnosis task requires a large number of tag data for training,and requires time and manpower to obtain enough labeled samples;At the same time,noise will interfere with the results of fault diagnosis.In order to overcome the above two shortcomings,this paper proposes a fault diagnosis framework based on semi-supervised learning,which extracts features from unmarked data and realizes fault classification by using contractive autoencoder and multi-layer perceptron respectively.The validity of the proposed framework is verified by gearbox data set and wind turbine benchmark model;Compared with the framework based on stacked autoencoder and stacked de-noising autoencoder,the proposed method has better fault diagnosis accuracy and robustness.Secondly,because some fault samples never appear or are difficult to obtain in fault diagnosis,the so-called zero sample fault diagnosis problem is caused,and the traditional deep learning method cannot achieve zero sample fault diagnosis.In order to realize two kinds of unseen fault diagnosis,this paper proposes a fault diagnosis framework based on anomaly detection and contractive stacked autoencoder,which is composed of feature extractor,zero-sample fault recognizer and fault classifier.The zero-sample fault recognizer is constructed based on the extracted features and the error between reconstruction and input by using the fault samples of the seen categories for training.Through the bearing data set of Case Western Reserve University,it is verified that this method can not only identify the visible and invisible fault samples in the data set,but also further identify the invisible fault samples through the designed discriminator,and at the same time identify the visible fault samples through supervised learning.Finally,aiming at the problem that there are many unseen classes in zero-sample fault diagnosis,especially the problem that there is no common attribute between unseen classes and seen classes,a multi-class zero-sample fault diagnosis method based on anomaly detection and comparative learning is proposed.In the fault detection part,the normal data is used for training,and the reconstruction error in the test phase is compared to determine whether there is a fault;In the part of fault identification of multiple unseen classes,the essential features of samples are extracted through comparative learning,and the anomaly detection is used for pre-classification,and different types of faults are determined through different spatial mapping.Compared with the method based on direct attribute prediction,the effectiveness and superiority of the proposed method are verified.