Monitoring Strategy Of Industrial Process Based On Data-driven

With the advent of the era of industrial modernization,the safety of the production process and the reliability of product quality have been paid more attention in recent years.The data-driven based process monitoring strategies have become a key technology in process industries for safety,quality,and operation efficiency enhancement.Therefore,the real-time monitoring of industrial process based on data-driven multivariate statistical analysis algorithms have become the focus in academic circles.Based on the research of many scholars,this paper discusses the problem of process monitoring with dynamic and multimodal characteristics.The main research contents of this paper are summarized as follows:(1)Aiming at monitoring multivariable dynamic processes with variables present significant features of time-series,a novel fault detection method based on independent component analysis-dissimilarity(ICA-Diss)is propose in this paper.Firstly,the independent components of the original high-dimensional data are extracted through independent component analysis(ICA).Secondly,two data sets in independent components are determined,one of which is composed of a default window and the other is the first nearest neighbor of the window data.After that,a new statistic is constructed to evaluate the dissimilarity of the two sets.Finally,the data is updated by gradually moving the default window,and the statistic of each window data is calculated to monitor the status of a dynamic process.In ICA-Diss,the latent information of the observation data can be captured by using ICA.meanwhile,the statistic based on independent components is capable of both reducing the influence of time-series on processes monitoring and improving the fault detection rate of dynamic processes.ICA-Diss is applied to a dynamic numerical case and the Tennessee Eastman(TE)process,and compared with the traditional principal component analysis(PCA),dynamic principal component analysis(DPCA)and ICA methods to verify the effectiveness of the proposed method.(2)A novel methodology based on independent component analysis-statistical characteristics(ICA-SC)is proposed to monitor multivariable systems whose variables present significant levels of autocorrelation.Initially,ICA is utilized to find the independent components of the process.Then,a window data set in the IC subspace is obtained using default parameters,such as window width and sliding size.Subsequently,the statistical characteristics of each window data set,including the mean and the variance of monitored independent components,are calculated.Finally,a new statistic index based on these calculated statistical characteristics is developed to monitor the status of the current process,and a fault diagnosis strategy based on the contribution charts of these independent components is used for analyzing the cause of abnormal change.The simulation results of a numerical case and Tennessee Eastman benchmark process illustrate that ICA-SC can reduce the computational complexity and improve the fault detection rate of a dynamic process with significant autocorrelation when compared with traditional multivariable monitoring methodologies such as dynamic principal component analysis(DPCA)and ICA.(3)In order to solve the fault detection problems of complex multi-modal process with variable scales and different degree of discreteness,a fault detection method based on locality preserving projections-neighborhood weighted mahalanobis distance(LPP-NWMD)is proposed.Firstly,the original high-dimensional data was reduced to low-dimensional space by means of locality preserving projections(LPP).Secondly,we determine the K-nearest neighbor set of the k-th nearest neighbor of each sample and calculate the weight of the sample in the low-dimensional space.Finally,we use the neighborhood weighted mahalanobis distances of samples as the statistics to monitor the quality of the process.The LPP-NWMD method reduces the computational complexity while maintaining the neighbor structure of raw data.In addition,the weight method eliminates the multi-modal and large variable scale characteristics of data and improves the fault detection rate of process.Compared with traditional PCA,k-nearest neighbor(k NN)and other methods,the results of numerical examples and semiconductor etching process verify the effectiveness of LPP-NWMD method.