Research On Improved Npe Algorithm For Batch Process Fault Detection

The batch process technology plays an important role in the production of highvalue-added products.Due to the increased automation of the production process,the production link becomes more and more complex,which puts forward higher requirements for the safe and reliable operation of the batch process.Therefore,to ensure production safety and product quality,the establishment of an effective monitoring model for batch processes,and timely and accurate fault detection and diagnosis have attracted much attention in the industrial field and academia.At present,the technical level of industrial production is constantly improving,and a large amount of production data is stored because these data contain rich process information.Therefore,it is necessary to use these process data to monitor the health of the process.How extracting the main feature information from a large amount of process data is the key to process monitoring.When traditional multivariate statistical methods are used for fault detection,it is usually assumed that the data has an ideal state and is not affected by outliers.However,the actual process data does not exist ideally,and due to the complex characteristics of batch process data,it is necessary to detect batch process faults on time and accurate detection is challenging.Therefore,based on the neighborhood preserving embedding algorithm,this thesis proposes a corresponding improved algorithm for the problem that the nonstationary,multi-modal,dynamic,and multi-stage characteristics of batch process data lead to poor fault detection results and effect.The main research contents are as follows:(1)In view of the mixed distribution characteristics of stationary and nonstationary data in the batch process,fault information is easily submerged in normal nonstationary signals,which makes fault detection difficult.A stationary-nonstationary Bayesian joint statistical indicator monitoring algorithm of variable grouping based on weighted orthogonal principal component analysis-exponential global neighborhood preserving embedding(WOPCA-EGNPE).First,the stationarity of the variables is tested by augmented dickey-fuller(ADF),the variables are grouped,and the stationary and the nonstationary space are divided;secondly,the nonstationary variables are obtained in the nonstationary space through the cointegration analysis method to obtain a stationary residual sequence,For stationary residual sequences,the exponential global neighborhood preserving embedding(EGNPE)algorithm is used to model and construct detection statistics,which algorithm can consider the global and local information of the data at the same time.The exponential transformation method highlights the important global and local data information;then,the weighted orthogonal principal component analysis(WOPCA)algorithm is used to model the detection statistics in the stationary space;finally,the Bayesian inference method is used to establish a joint detection index,Realize the monitoring of nonstationary batch processes.The validity of the detection is verified by the simulation results in the penicillin fermentation process.(2)To solve the problem of fault detection caused by the coexistence of multimodal and dynamic characteristics of batch process data,a sparse weighted neighborhood preserving embedding(SWNPE)algorithm based on weighted double neighborhood standardization(WDNS)is proposed.Firstly,based on finding the double nearest neighbors of the sample,the weighted double nearest neighbor set is obtained,and the weighted double nearest neighbor set information is used to standardize the sample,and the multi-modal data is processed into a single modal distribution,which eliminates the difference of multi-modal center points and effectively solves the multimodal characteristics.Secondly,considering that the NPE algorithm cannot better deal with the problems caused by the dynamic characteristics,the SWNPE algorithm is obtained based on the NPE algorithm by using the inverse distance weighting and local optimal sparse representation.The SWNPE algorithm not only deals with the dynamic characteristics of the data,but also enhances the robustness of noise and outliers.Finally,the weighted double nearest neighbor standardized SWNPE model is used to realize fault monitoring.By comparing dynamic neighborhood preserving embedding(DNPE),temporal neighborhood preserving embedding(TNPE),sparse weighted neighborhood preserving embedding(SWNPE)and local neighbor normalizationneighborhood preserving embedding(LNS-NPE)in the simulation process of penicillin fermentation shows that the proposed WDNS-SWNPE algorithm has a higher detection rate and improves the dynamic and multimodality the ability to monitor batch processes with characteristics coexisting.(3)Due to the multi-stage characteristics of batch process data,the overall modeling cannot reflect the actual working conditions of the batch process,which will lead to inconsistency between the detection results and the actual process.In addition,the detection process based on the NPE algorithm maintains the local geometric structure of the data while ignoring the global information,while the latent variables extracted from the global or local information of the data alone cannot fully characterize the process data.Therefore,a multi-stage optimization regularized neighborhood preserving embedding(ORNPE)algorithm is proposed.Firstly,the dynamic global information is extracted by the slow feature analysis algorithm and local structural information through a neighborhood preserving embedding algorithm.Then,crossentropy is introduced to optimize the global dynamic information;finally,affinity propagation(AP)algorithm is used to divide multiple process stages.The experimental results of penicillin fermentation simulation process and semiconductor etching process verify that the proposed algorithm has better detection effect.