The Research On Batch Process Fault Detection Based On Neighborhood Preserving Embedding Algorithm

Nowadays,due to the high requirements and strict standards for product quality,production efficiency and safety performance,fault detection has also received widespread attention from scientists with the rapid development of computer technology and artificial intelligence.More and more researchers are devoted to improve the accuracy and efficiency of fault detection,and then by using certain theories,actual industrial processes are guided and helped.The traditional multivariate statistical methods based on data-driven need to assume that fault detection is not affected by noise,outliers,etc,and process data have approximately linear state and single working condition,which is beneficial to subsequent modeling and analysis.However,the increasingly complex and intelligent industrial processes are no longer suitable for directly utilizing the above-mentioned traditional methods,so it is urgent to improve them to obtain better results.Aiming at the complex characteristics of non-linearity,non-Gaussian,dynamic,multi-stage,and multi-mode mixed distributions in batch processes,which often have negative impacts on the efficiency and accuracy of fault detection,this thesis proposes several improved algorithms for fault detection by analyzing the structural characteristics of Neighborhood Preserving Embedding(NPE)algorithm and combines the characteristics of specific batch processe.The research contents are as follows:(1)Aiming at the characteristics of the multi-mode of batch process data and the large differences in each modal structure,Gaussian and non-Gaussian mixed distributions,an LNSNPE-SVDD fault detection algorithm based on Local Neighbor Normalization(LNS)is proposed.Firstly,the local neighbor set of the original data is seeked and the data of local neighbor set are standardized.At the same time,the multi-mode data are integrated as a single mode and process the coexisting data of Gaussian and non-Gaussian are processed as the approximate multivariate Gaussian distribution.Then,the dimensionality reduction is performed while maintaining the local data manifold structure is preserved effectively and the local features are extracted fully.Finally,the fault detection model is established by using Support Vector Data Description(SVDD),and monitoring statistics are constructed for process monitoring,which can further improve the detection rate.(2)In order to solve the problem that the Neighborhood Preserving Embedding algorithm only pays attention to the local structure information of the data and doesnot consider the global information,which causes low detection rate for the fault detection of batch processes with complex dynamic characteristics,an improved NPE algorithm based on Cross Entropy(CEGLNPE)is proposed,which can take into account both global and local data utilization and improves the efficiency and accuracy of fault detection.Firstly,Cross Entropy algorithm seeks the global optimum through multiple iterations of data by updating the probability density,while NPE algorithm maintains the local structure.Then,CE and NPE algorithms maintain the global structure and the local structure,respectively,the new objective function is structured.Finally,the sliding window is used to update the data to solve the dynamic problem and establish a global-local fault detection model.Comparing with KPCA and NPE algorithms,the results of the artificial data set Swiss-Roll and penicillin fermentation simulation process verify that the proposed algorithm has better effectiveness.(3)In view of the multi-stage characteristics of batch process,that is,the data structure between the stages is different,when a process is modeled as a whole,the differences between some structures would be ignored,and this may cause poor effect of fault detection.A multi-stage NPE algorithm based on Sparse Subspace Clustering(SSC)is proposed for fault detection.Firstly,k-nearest neighbor is introduced as the constraint term of SSC,and the global and local aspects are taken into account to divide the data into clustering stage.Secondly,NPE algorithm is used to reduce dimensions and extract the features for each sub-stage.Thirdly,the wavelet transform is used to reduce the noise the statistics which can eliminate the influence of noise and interference for the results,this can minimize the loss rate of data.Finally,the results of the penicillin fermentation simulation experimental platform verify the effectiveness of the proposed algorithm.