Research On Bidirectional Fault Diagnosis Method Based On Heterogeneous Ensemble Incrementai Mergence

With the advancement and development of the Industrial Internet.Massive data generated by running equipment is continuously collected,transmitted,and stored.How to use current intelligent algorithms to efficiently analyze fault information,which become a hot issue in intelligent manufacturing today.Meanwhile,it can not only premise the accuracy,but also shorten the diagnosis time and use the storage space efficiently.At present,there are still some problems in fault diagnosis: facing massive data,because the classification model is difficult to balance stability and plasticity,how to explore a classification model that can be adjusted in real time according to the state and attributes of the equipment,and mine the potential information of new data is important.Although incremental learning can retain the knowledge that is already learned,and continuously learn the information in the new data.However,in the process of fault diagnosis,the data stream generated from incrementally has the characteristics of mass,imbalance,stability-plasticity and difficulty in balance.Therefore,this study considers the characteristics of industrial data and the diversity of fault information,and then constructs a bidirectional fault diagnosis model based on heterogeneous ensemble incremental mergence,which realize the fault identification and diagnosis based on massive unbalanced high-noise rolling bearing equipment status data.Main works are as follows:First,this paper proposes a dynamic feature ranking method based on extreme gradient boosting of mechanical equipment fault data(XGBDFR).This method sort feature importance values according to the gradient boosting principle.With the addition of incremental data,the feature importance sequence is dynamically adjusted in real time;at the same time,the candidate set is updated in real time through reverse verification accuracy,which is used to solve the problem of feature filtering in the ensemble classifier to process of massive new data efficiently.Then,in order to solve the unbalanced problem that is common to fault data,this paper introduces Introducing resampling non-equilibrium processing method based on partitioned safety set which divides the safe set and non-safe set through the calculation of k-safety.Using BSL-SMOTE oversampling method for minority samples in unsafe sets,and using CBNM method for the majority of samples in the safety set.After that,in order to solve the problem that the plasticity-stability of the classifier is difficult to balance,this paper create a bidirectional weight adjustment mechanism for heterogeneous base classifiers,which improve the weight of each base classifier simultaneously in the training data and the prediction data through an improved locally sensitive hash algorithm.In this way,it can not only make the ensemble classifier take into account historical data to maintain classification stability,but also enable the base classifier to make adaptive adjustments based on current prediction data,which improve plasticity of classification models.Finally,through training and testing incremental data to make adjustment of heterogeneous classification models,a bidirectional fault diagnosis model based on heterogeneous ensemble incremental mergence(HEIMB)is formed.Using this method for bearing equipment data for experiments,the results show that this method makes the bearing fault diagnosis efficiency reach 89.74% on average,4.32% improvement over SVM,XGBOOST,and CDAE methods without incremental learning on average,3.84% average improvement compared with other non-equilibrium data processing methods,and 2.40%improvement compared to XGBDFR method.This method can realize the identification and diagnosis of fault categories based on massive unbalanced strong noise rolling bearing equipment data.