Research On Fault Self-Learning Method Of Rolling Bearing Based On Two-Level DBN

Rolling bearing is an important part of rotating machinery.Its working condition directly affects the working performance of the whole equipment.The operating environment of the bearing is complex and it is easy to cause failure.Bearing failures often cause serious accidents.Therefore,It is a crucial issue in engineering applications for accurate fault identification and classification of rolling bearings.With the rapid development of modern industry,more mechanical equipment has been added to industrial production,bearing failures occur frequently,and mechanical failures exhibit a "big data" characteristic.Other shallow models such as BP neural network are faced with problems such as dimensional disasters,and the fault data acquired on site is often incomplete and unlabeled.Therefore,it is difficult to establish an effective diagnostic model for traditional pattern recognition methods.There is an urgent need for self-learning diagnostic algorithms and model development and research,which has become one of the main research hot spots and urgent problems to be solved.Deep Belief Network(DBN)adopts unsupervised layer-by-layer greedy training method to avoid manual operation of feature extraction and selection.It has the ability to process high-dimensional and nonlinear data,and can effectively prevent dimensional disasters.It is very suitable for dealing with the fault diagnosis problems of industrial ”big data” in the new era.Previous studies have shown that a significant characteristic of DBN is that it can directly derive from high-level features by layer-by-layer greedy learning from the low-level original signal.Therefore,this passage based on the raw data of rolling bearings,combined with S-transformation and Bayesian classifier,a self-learning method for rolling bearing faults based on two-level deep confidence network(DBN)is proposed.The main research contents are as follows:(1)Establish a DBN network with multiple hidden layers.In the case that fault type is complete and the fault level is incomplete,the original bearing fault data is normalized as the input of the DBN,through the feature extraction of the DBN and the classification of the classifier,the fault type is accurately identified,and compared with the classification effect of BP neural network,the results show that DBN has the feature extraction ability unmatched by other shallow networks.(2)Using the powerful feature extraction capabilities of DBN,establish a two-grade DBN of fault self-learning model for bearings.First,establish DBN1 fault typeself-learning model.After performing S transform processing on the original bearing fault data,the S transform matrix is obtained.the result is a two-dimensional matrix with rows representing frequency and columns representing time.By extracting the amplitude mean information of the time column as the input of DBN1 for extracting the feature,the extracted features are used as the input of Bayesian classifier.The self-learning of the fault type classification model is realized by dividing the posterior probability output of the Bayesian classifier into reasonable confidence intervals belonging to the same fault type or different fault types.Secondly,the DBN2 fault damage level self-learning model is established,and the fault level is classified for the data that has been identified by the fault type,and the data is normalized and used as the input of the DBN2 to extract features.and the extracted features are used as input of Bayesian classifier,select the same confidence interval division method as the fault type classification to implement self-learning of the fault damage level classification model.Finally,the two-level DBN model is used to realize the intelligent classification of bearing fault types and damage levels.