| Safety and quality are two of the most important goals of modern industrial production.Quality related process monitoring technology is the key to achieve these two goals.In recent years,advances in sensors and computer technology greatly enriched the available process data.This promoted the rapid development of Multivariate Statistical Process Monitoring(MSPM).The application of traditional MSPM method has many limitations.For example,the process data must obey the Gaussian distribution,must obey linear distribution and not contain non-stationary variables,etc.This paper builds quality-related monitoring models based on Slow Feature Analysis(SFA)method,and proposes a new fault detection method for the following situations,including:(1)In order to overcome the shortcomings that previous quality-dependent MSPM method cannot correctly distinguish the change of operating condition from the real fault,a quality-related dynamic slow feature analysis fault detection method was proposed.This method combines Canonical Correlation Analysis(CCA)with Dynamic Slow Feature Analysis(DSFA)to construct a new optimization objective function.This method can extract slow features strongly related to quality in process data.Process data are decomposed into quality-related subspaces and quality-unrelated subspaces.Then,establish steady-state and dynamic monitoring indexes respectively.This method uses the process data obtained in real time to detect faults and determine whether they affect the quality.It can distinguish the deviation of normal working conditions from the real faults and has good dynamic performance.Finally,Tennessee Eastman(TE)process simulation is used to verify the effectiveness of the proposed method.(2)In order to overcome the shortcomings that previous quality-related MSPM method is not suitable for non-stationary processes,this paper proposes a new method.This method first identifies the non-stationary process variables and stationary process variables in the system.Then uses Gonzalo-Granger decomposition to solve the common trend model,so as to separate the stationary part and non-stationary part of the non-stationary data.Then,integrate stationary data and stationary subspace of non-stationary data.And a quality-related monitoring model is established by using DSFA and CCA to realize effective monitoring of non-stationary quality variables.Finally,numerical simulation and three-phase flow simulation experiments are used to compare the previous methods.It is proved that the proposed method can accurately detect the quality-related faults in non-stationary systems.(3)In order to overcome the shortcomings that linear mass dependent slow feature MSPM method is not suitable for nonlinear process data,this paper proposes a new method.In this method,Kernel Dynamic Slow Feature Analysis(KDSFA)and Kernel Canonical Correlation Analysis(KCCA)are combined to calculate the steady-state and dynamic monitoring statistics reflecting the quality changes.The process changes are comprehensively analyzed by combining the global monitoring index.Finally,a numerical simulation and Continuous Stirred Tank Reactor Process(CSTR)simulation were conducted to verify the reliability and effectiveness of the quality-related fault detection method based on nuclear slow characteristic analysis. |
