Fault Identification And Location For Industry Process Based On Ensemble Models

With the rapid development of industrial processes in the direction of complexity,automation,and large-scale,the integration of industrial systems is getting higher and the structure of each unit in the system is also becoming more complicated.For such a large system,once a device fails,it is highly likely that the entire system will break down.Therefore,improving the stability and safety of industrial process has become a focus of attention,and fault diagnosis technology is an effective way to solve this problem.The diagnosis method based on data and statistical machine learning has received extensive attention in recent years.Ensemble learning is a machine learning enhancement algorithm.It can be used to integrate classification models and improve the accuracy of recognition.It is well suited to apply ensemble learning into the field of fault diagnosis.Based on the above considerations,this paper selects SOFC system as the simulation platform,and studies the fault identification and localization of industrial process.The proposed combination of ensemble learning and statistical machine learning can improve the monitoring effect of industrial process.The main research work of this paper includes:1.For the high-dimensional,nonlinear,complex distribution and variable working condition of data during system aging,a multi-fault identification strategy of PCA-HSVM is proposed from the perspective of pattern classification.PCA is used to extract fault features in the origin high-dimensional data space to reduce noise.HSVM is constructed according to the hierarchical integration strategy using the classification accuracy on the validation set.Compared with traditional multi-classification SVM based on one-v-one or one-v-rest strategy,HSVM requires fewer classifiers for training while guaranteeing recognition accuracy,and effectively avoids the problem of training samples imbalance.This method has achieved good fault identification on the simulation platform of SOFC system.2.According to the difficulty of a large quantity of concurrent fault modes and the challenge of concurrent fault identification,this paper uses multi-label learning to solve the problem of concurrent fault identification.Since the concurrent faults are generated by the mixture of single faults,the concurrent faults can be described by vectorized labels,which builds the internal relationship between single faults and concurrent faults.Then the ensemble classification model GBDT is parallelized called PGBDT,and each sub-GBDT outputs a binary label indicating whether a corresponding single fault occurs.The advantage of this method is that only n-dimensional binary label can represent the 2~n concurrent fault types,meanwhile PGBDT also only needs single fault data as training set.Compared with the traditional supervised model based on scalar-label,the proposed method has better identification performance on the SOFC systems.3.In order to solve the fault location of industrial process,DBSCAN-RF,a fault location algorithm based on feature clustering and feature selection is proposed from data-driven perspective.The DBSCAN clustering method is used to group the highly correlated features into a group,and combined with the RF feature selection algorithm,the features that are more important for fault classification are selected from each cluster to obtain the final result.DBSCAN-RF enables the removal of redundant features while picking features that are closely related to fault occurrence.The simulation results on the SOFC system suggests that new method obtains more accurate location effect than RF feature selection method.