Application Research Of Machine Learning In Chemical Process Fault Detection

With the continuous progress of industrial level,chemical process is developing towards large-scale and complicated direction.This means that the risk of failure is greatly increased.If a fault occurs in a certain link,it is easy to lead to a chain effect in the whole process,and serious cases will lead to major accidents.Therefore,fault detection is particularly important.In recent years,with the progress of instrumentation technology and data storage technology,massive production data is easy to be retained,which makes people pay more and more attention to data mining and analysis,and data-driven machine learning methods are applied to the monitoring of industrial processes.However,due to the complexity of the system and the defects of the algorithm itself,the commonly used fault detection methods are not ideal when dealing with nonlinear process and early minor faults,and the accuracy of fault identification is not high.In order to solve the above problems,this paper studies the method of chemical process fault detection based on machine learning.Firstly,a new feature extraction method MIPCA was proposed by combining Mutual information with principal component analysis(PCA).The mutual information matrix was obtained by calculating the mutual information among variables,and the feature extraction was carried out instead of the covariance matrix in principal component analysis(PCA).Both linear information and nonlinear information were retained,which improved the effect of principal component analysis on nonlinear data processing and improved the processing ability of nonlinear data,and the effect was better in the face of chemical process data.Secondly,a small amount of fault data is introduced into the SVDD algorithm,the control limit calculation method is improved and the particle swarm optimization algorithm(PSO)is used to optimize the parameters,and a new fault detection algorithm CSVDD is proposed.The classical SVDD algorithm only uses normal data as the training set,the control limit accuracy is not high,and the fault detection rate is low when encountering minor faults or initial faults.The CSVDD algorithm improves the accuracy of the control limit and significantly improves the fault detection rate,especially the small fault.PSO algorithm is used to solve the problem that parameter debugging is difficult and the model is easy to fall into local optimal.Then,combining MIPCA and CSVDD,a new chemical process monitoring method,MIPCA-CSVDD,was proposed.It includes two parts: offline modeling and online detection.In the off-line modeling part,the original data is first standardized to solve the problem of dimensional disunity.Feature extraction is carried out on the original data through MIPCA,and then the CSVDD algorithm is used to establish a fault detection model to determine whether the fault occurs.In the online detection part,the new data is standardized,and then feature extraction is carried out through MIPCA.Finally,the control limit of CSVDD model is used to judge whether the data is abnormal,so as to achieve the purpose of fault detection.Finally,the concept of Kernel Density Estimation is incorporated into the multi-SVDD fault identification algorithm for fault identification,which is called KD-SVDD fault identification algorithm.The accuracy of the SVDD model is very low for the data in the overlapping region.To solve this problem,kernel density estimation is introduced.When the data to be measured is located in multiple hyperspheres,the relative density of each sample among its similar samples is calculated.These samples are located in the overlapping region of multiple hyperspheres.Then,it is classified according to the mean of kernel density of various samples.The accuracy of multi-class fault identification algorithm is improved effectively.