Fault Detection And Diagnosis Of Batch Process Based On MLLE-OCSVM

With the rapid development of modern information science and technology,more and more new technologies are widely used in modern industrial production process.In recent years,owing to the increasing variety of chemical products and more and more diversified production demand.The batch process is in line with the production requirements of high value-added,small batch and variety.It has been widely used in the production of pharmaceuticals,cosmetics and biological products in the process.With the scale and the increasing complexity in the production process,the requirements of the safety and reliability are getting higher and higher.Therefore,it is very important to successfully monitor the batch process.In the background of the batch process,the traditional multivariate statistical algorithm needs to assume that the process variables obey the Gaussian distribution requirement,and the constructed monitoring statistics that based on the Mahalanobis distance cannot obtain the structural information of the active component,leading to the problem of reducing the failure detection the rate.We research a new method based on the local linear embedding method(Locally Linear Embedding,LLE)and the one-class support vector machine(OCSVM),the following methods are studied:(1)A multi-directional Locally Linear Embedding(MLLE)method is proposed and applied to the feature extraction of batch process.In the batch process,the traditional multivariate statistical methods have the prerequisite,such as assuming that the monitoring variables are linear,the process parameters do not change with time and so on.However,in practical industrial processes,industrial objects are difficult to meet the above assumptions,and the improved algorithms currently used in the batch process,most of which do not fully consider the complex characteristics of complex industrial processes,they are only on the basis some features of traditional multivariate statistical methods to improve,cannot get a good feature extraction effect.In this paper,the local linear embedding method in the manifold learning algorithm is applied to the feature extraction process of the batch process,and the control limit is determined by the traditional statistics.Finally,the experiment was carried out on the simulation platform of penicillin fermentation.The experimental results show that the method can monitor the batch production process well.(2)In-depth study based on multiway local linear embedding(MLLE)method,and improve its fault monitoring method in the batch process.On the basis of the research(1),the two core parameters(neighborhood parameters and dimensionality dimension)of the method are optimized for the unique characteristics of the batch process,and the adaptive ability and monitoring performance of the monitoring system are improved.Because of the MLLE method,the neighborhood parameters and dimensionality dimension are needed before the feature extraction.However,in the process of using this method,most of the parameters are set by the experience,and exits the unstable factors;and then,collecting data of each time are not the same in the process,they are random and fluctuant,there is poor reliability simply rely on the experience to set parameters.In this paper,the adaptive neighborhood selection method and maximum likelihood estimation method are adopted for neighborhood parameters and dimension parameters.After optimization,the MLLE method was used to verify the experiment in the penicillin fermentation simulation system.The results show that the parameter optimization method can improve the performance of the feature extraction and batch process.(3)For improving the performance of monitoring statistic,a fault monitoring method for batch process based on multi-directional local linear embedding method(MLLE)and One-class support vector machine(OCSVM)is proposed.The traditional multivariate statistical monitoring methods itself have some assumptions,such as monitoring variables subject to independent distribution,etc.,and in the application of it to the fault monitoring,made additional assumptions,such as the system input and output were linear correlation,variable variables obey Gaussian distribution and so on.However,in the actual industrial production process,these ideal conditions are often difficult to meet,and the constructed monitoring statistics based on the Mahalanobis distance will lead to failure detection rate.Aiming at two problems,this paper combines the one-class support vector machine(OCSVM)with linear embedding method(MLLE),uses the MLLE to extract the data of the process data,and then uses the OCSVM to train its data samples,then fault monitoring after constructing the nonlinearity statistics.(4)In this paper,the batch process fault monitoring method proposed is used to validate the data collected by industrial field.In this paper,using the fault monitoring method based on the actual experimental data of Escherichia coli fermentation provided by a biological pharmaceutical factory in Yizhuang,Beijing,studied.The results show that the proposed monitoring method can guide the actual production process well,can monitor the fault in time,and have practical value for the actual production.