Operating Performance Assessment Of Complex Nonlinear Industrial Processes Based On Local Information Preservation

Due to the aging of production equipment,uncertainties and disturbances,the industrial processes are easily deviated from the optimal performance and cannot guarantee the optimal product quality and comprehensive economic benefits,so it is necessary to further measure the gap between the current operating performance and the optimal operating condition on the basis of normal industrial process operation,judge the non-optimal operating performance,identify the causes leading to nonoptimal operation,and ensure that the industrial process operates in the optimal performance.However,as the scale and complexity of the process industry increases,nonlinearities in the data are prevalent,and direct assessment models based on linear assumptions fail to represent the nonlinear correlations between the original variables,resulting in limited performance of the assessment models.In addition,actual industrial processes often have nonlinearities along with one or more other characteristics,such as dynamics,strong autocorrelation and intercorrelation.In complex industrial processes,modeling properties other than nonlinearity is often more complex and challenging than modeling only nonlinear properties.Thus,in this thesis,the study of operational performance assessment is carried out for input-output related strong nonlinearity,dynamic nonlinearity,and higher-order dynamic nonlinearity,respectively,using local information preservation techniques under the premise that the process has nonlinear characteristics,and the detailed work is as follows:(1)To address the problem that the process is strongly nonlinear in the actual industrial process and many existing operating performance assessment methods for studying nonlinear problems ignore the local neighborhood information of each performance grade,resulting in poor accuracy of the assessment results,a kernel locally linear embedding partial least squares(KLLEPLS)method is proposed and applied to the operating performance assessment of complex industrial processes.In the offline modeling part,the local information extracted from the kernel locally linear embedding(KLLE)is embedded into partial least squares(PLS),and the KLLEPLS algorithm is proposed,which has the ability of PLS algorithm to maximize the correlation between process variables and comprehensive economic indexes(CEI),and the ability of KLLE algorithm to maintain the local nonlinear structure;in the online assessment part,the assessment is based on similarity assessment index;when the assessment result is nonoptimal,the contribution rate of variables is calculated according to the local and global information extracted online to identify the causes of non-optimal.Finally,the effectiveness of the algorithm is verified by the dense medium coal preparation process.(2)For the dynamic nonlinear problems existing in the process,the operating performance assessment method of complex industrial processes with weighted supervised probabilistic slow feature analysis(WSPSFA)is proposed.In order to apply to the complex dynamic nonlinear process,two similarity strategies are introduced to assign different weights to the objective function of SPSFA: one represents the similarity in static information,which is derived by the similarity between the training samples and the online samples,and the other is the similarity in dynamic information,which is expressed by the first-order time difference of the training samples and the similarity of the first-order time difference of the online samples;then,the weighted log-likelihood function of WSPSFA is constructed using these two weights,and the optimal model parameter set is obtained using Expectation Maximum(EM)algorithm;in order to describe the process performance more carefully,two assessment indexes are constructed to capture the static and dynamic changes comprehensively;for the nonoptimal performance of the operation,the algorithm of sparse learning is used to achieve the accurate tracing of the non-optimal causes.Finally,the effectiveness of the proposed algorithm is verified by the dense medium coal preparation process.(3)In the above study,considering that different variables have strong autocorrelation and intercorrelation and the process exhibits high-order dynamic characteristics,a method for assessing the operating performance of complex industrial processes based on weighted autoregressive supervised probabilistic slow feature analysis(WAR-SPSFA)is proposed.Firstly,the first-order autoregressive model in the slow feature analysis is extended to a higher-order autoregressive model,and the higher-order state space equations are formed in the probabilistic framework to extract the slow features with higher-order time correlation to capture the higher-order dynamic information in the process data;the comprehensive economic information is incorporated into the assessment model to extract the feature information related to the comprehensive economic indexes;the performance transfer function is treated with the observed values by local linearization nonlinearities in the emissivity function to improve the processing capability of complex nonlinear processes;and identify possible causes based on sparse learning methods when non-optimal operating performance occur.Finally,the effectiveness of the algorithm is verified in Tennessee Eastman.