Research On Data-driven Dynamic Process Fault Detection Method

With the advancement of social science and technology,modern industrial processes have continued to develop into large-scale and complex,and the causes and frequency of failures have also increased.Whether it is the safety of human life or the company’s property,the consequences are irreversible,so it is essential to timely and accurately diagnose faults and predict the occurrence of faults.Due to the complexity of data structures,data-driven methods are widely used as an essential foundation in fault diagnosis.There is no need to rely on an accurate model;only a large amount of data collected from production equipment is used as a basis to deal with complex characteristics such as nonlinearity,Gaussian,time-varying,and dynamic existing in the process data set.Traditional methods based on data-driven often use a single model modeling method to ignore the local information of the process.Using online measurement data for static monitoring results in the inability to adapt to the actual industrial process and fail to detect the failure in time and effectively.To this end,based on the slow feature analysis(SFA)method,the paper uses the data generated by the Tennessee-Eastman(TE)process as the research background to detect faults in dynamic industrial processes.The main contents of the paper include:(1)To solve the problem that local information is often ignored by a single modeling method,traditional methods ignore the dynamic characteristics of actual industrial processes.A fault monitoring and diagnosis method were proposed based on multi-block just-in-time learning slow feature analysis(JITL-MBSFA).Use the SFA algorithm to simulate the failure of the static and dynamic process.By dynamic expansion of the sample,the front and rear sample data are no longer independent.Introduce Mutual Information(MI)algorithm to extract variable modules with the most extensive similar features so that the monitoring process retains all characteristic information.It is then divided into two sub-blocks based on the correlation between variables.Introduce a just-in-time learning algorithm(JITL)and screen the most related feature information from variables containing all features as the optimal training set.Then train offline models through local modeling methods.Finally,the effectiveness of the method was verified in the TE experiment.(2)According to the traditional method of partitioning molecular blocks according to experience,the feature information in variables can not be fully utilized and thus affect the detection results.The model obtained offline from historical data cannot adapt to the time-varying characteristics of the process.A multi-block sliding window slow feature analysis(KL-MWSFA)fault diagnosis method based on KL divergence was proposed.Based on dynamic data expansion,the difference between sub-blocks and whether the feature information of process data can be fully utilized will affect the final detection effect,so the KL divergence algorithm is introduced.KL divergence component was used to represent the similarity of features between variables.The initial centers of variables were determined to iterate and update the central sample points to realize unsupervised partition of sub-blocks and avoid the error of artificial partition of molecular blocks.In the process of local modeling,the sliding window algorithm is introduced to enhance the adaptive ability of the model to time-varying characteristics.Finally,the simulation results in TE show that this method can effectively improve the fault detection ability.