Industrial Process Modeling And Fault Detection Based On Trajectory Analysis

In modern Industry,process monitoring technologies play an important role due to its value and significance for research and practical use.Since industrial processes are becoming more and more complicated and large-scale,the mechanism model of processes are difficult to obtain,while the development of computer technologies and distributed control system helps to collect a large amount of process data and the big data consists of crucial information related to processes.Therefore,scholars focus on extracting process features and valid information from process data and further achieve process modeling and online monitoring based on massive process data.In past decades,various progresses have been made to meet different demands of process monitoring in practical industry in the field of multivariate statistical process monitoring(MSPM).However,traditional MSPM methods are under ideal assumptions and limited to industrial applications.To deal with complicated process natures of industrial big data such as dynamics,nonlinearity,non-Gaussian property and stochastics,this thesis focuses on process dynamics and solves a series of other problems by analyzing process trajectories for batch processes and continuous processes.(1)A fault detection method based on trajectory analysis is proposed to handle with process dynamics and uneven length problem of batch data for batch processes.After constructing trajectory vectors and extracting similar trajectories by just-in-time learning algorithm,monitoring models are obtained for online fault detection to improve the performance of dynamic process monitoring with uneven batch length problem.In application,due to the advantages of trajectory vectors which emphasize more on process changes and trends,reliable monitoring results can be achieved.Meanwhile,the disturbance caused by uneven length of batch data and missing data problems can be avoided.(2)A quality-relevant fault detection method based on local trajectory analysis and PLS-SVDD algorithm for between-phase transitions of batch processes is proposed to deal with the complicated data features during this period.Except for strong dynamics,data during the transitions is more complicated compared to other periods,presenting non-Gaussian feature and nonlinearity as well.Meanwhile,process quality in the transitions shows important significance to indicate the process conditions in the former and latter phases,which means the quality-relevant monitoring for transitions is necessary.To accomplish this scheme,a local PLS-SVDD model is constructed to implement quality-relevant fault detection by regression model and overcome the disturbance caused by data characteristics such as non-Gaussian feature.During the fault detection for the whole procedure of batch processes,individual modeling and monitoring during transitions can provide a more reliable performance.(3)A fault detection method based on stochastic programming and trajectory analysis is proposed to handle with process stochastics and uncertainty during batch process monitoring.Taken process dynamics and quality-relevant monitoring into consideration,this method calculates the optimal history quality trajectory as the reference of fault detection with stochastic parameters on the basis of stochastic programming.Meanwhile,since the historical data collected by sensors may have missing data and limited batch problems,bagging algorithm is introduced to resample data and construct several sub-models with individual optimal quality trajectories,while corresponding monitoring statistics are designed.For online monitoring,by combining fault detection results of each sub-model with ensemble learning strategy as decision fusion,online quality-relevant fault detection is implemented.By introducing stochastic programming,this method eliminates the disturbance caused by process stochastics effectively after solves regular problems such as process dynamics and quality-relevant fault detection and finally shows better monitoring performance and application value compared to traditional methods.Finally,the fault detection method based on trajectory analysis proposed in this thesis is combined with practical industrial continuous processes to evaluate the performance.The results illustrate that the proposed method provides satisfying monitoring performance with excellent application value.