Research On KNN Fault Detection Method Under Multi-block Modeling Strategy

Modern industrial processes are becoming increasingly complex and fault types are increasing.Once a fault occurs,it will not only reduce economic benefits,but also cause serious casualties.With the development of computer technology and data acquisition equipment,some industrial processes have accumulated rich process data,which makes the data-driven fault detection technology continue to progress and become an important key technology to ensure the safe operation of the industry.Multi-block modeling strategy can effectively monitor largescale and complex industrial processes by establishing multiple sub-block detection models for the whole industrial process,which has certain advantages compared with the global model.In this paper,the improvement of fault detection method based on k-Nearest Neighbor algorithm is studied under multi-block modeling strategy.The main contents are as follows :(1)Aiming at the problem that the traditional k NN-based fault detection method does not consider the local information of the process and only establishes a global model,resulting in the low alarm rate,a multi-block k NN fault detection method based on mutual information is proposed.This method uses mutual information between variables to construct sub-blocks to extract local information of the process,so that the variables in the sub-blocks have more same information.On this basis,the fault detection model based on k NN is established for each variable sub-block,and the Bayesian inference method is used to integrate the detection results of each sub-block,so that the overall detection effect is more intuitive.Further,the fault diagnosis method based on Mahalanobis distance is adopted.By calculating the Mahalanobis distance between each variable and its mean value in the sample,the source variable causing the fault is found and isolated.Finally,the Tennessee-Eastman(TE)process simulation experiment verifies that this method has higher alarm rate than the traditional fault detection method.(2)Aiming at the problem that the fault detection method based on variable block has a low alarm rate for faults such as small offset and pulse oscillation,a multi-block k NN fault detection method based on two-layer information extraction is proposed.The first layer uses canonical correlation analysis to calculate the correlation coefficient between variables to construct variable sub-blocks to obtain local information;the second layer extracts observation information,cumulative information and the rate of change information data from each variable sub-block as information sub-blocks,and amplify fault differences by extracting feature information.Considering the distribution characteristics of data and the sparse degree of local area samples,the k NN fault detection model based on Mahalanobis distance is established for each information sub-block,and the detection results of all sub-blocks are fused by Bayesian inference method to integrate the advantages of each sub-block and improve the overall detection performance.(3)Aiming at the problem that abnormal information causing faults in the traditional k NNbased fault detection method is easily overwhelmed by normal information,resulting in untimely fault detection and a low alarm rate,a k NN fault detection method based on reconstruction error is proposed by using the auto-encoder and multi-block modeling strategy.In this method,the auto-encoder model is trained using the normal working condition data set,and the reconstruction error is extracted based on this model to solve the problem that the abnormal information is easily submerged.Further considering the fault characteristics such as small offset and oscillation,the multi-block modeling strategy is adopted to calculate the statistics of each sub-block and fuse the detection.The feasibility and superiority of the proposed method are verified by numerical simulation and TE process simulation experiments.