Methodologies Of Dynamic Time Delay Mining For Industrial Process Variables With Applications

In traditional data driven industrial process monitoring methods,time delays information among process variables are rarely considered,or simply using constants to describe them,which no doubt definitely affect the effectiveness of process monitoring models and the accuracy of monitoring knowledge representations,furthermore leading to the inaccurate and ineffective implementation of process monitoring tasks.Therefore,it is of a significant importance both in academic researches and engineering applications to explore effective methodologies for extracting dynamic time delays among process variables so as to represent and reason more accurately for process monitoring knowledge.This thesis proposes a dynamic time delay analysis(DTA)method which can effectively estimate dynamic time-delay information between process variables.Moreover,a Dynamic Time Fuzzy Petri Nets(DTFPNs)is presented,which can effectively describe transfer of dynamic information between process variables.The conditional distribution colored graph is used to express the dynamic time delay information before the activeness and dynamic reachability analysis method of dynamic time fuzzy Petri net is given.Furthermore,an effective process monitoring knowledge representation model is established along with the reasoning method.The main contributions of the thesis are presented as follows:1.Dynamic characteristics of time delays between process correlated variables are major concerns in the process control community.In response to this problem,this thesis proposes a dynamic time delay analysis(DTA)based on the technology of time series data mining,aiming at effectively estimating transfer time delays between process correlated variables.Employing dynamic sliding windows,dynamic time delays can be estimated offline by calculating similarities between correlated variables.2.The time delay is approximately estimated by static sliding time windows,which could not better deal with the dynamics of time delays.In response to this problem,we proposed a dynamic time delay analysis(e-DTA,dynamic time delay analysis by elastic windows)method based on similarity elastic windows,which is aiming at effectively estimating the transfer time delay between process variables.According to contrasting similarities between correlated variables,the size of the elastic window is self-tuned and the dynamic delay time can be estimated offline.3.In order to realize online estimation of dynamic delay information,two improved time series prediction methods are proposed:fuzzy interpolation time series prediction method and dynamic time delay prediction method based on deep learning network CNN.In order to solve the limitation of the traditional time series prediction,which is difficult to adapt to the mutation signal.The correlation analysis of time series delay variables and the process variables,the main effect of a single variable time delay correlation information and a plurality of main related variables can be obtained respectively,further introducing the time series prediction model,the online estimation process between variables related to transfer delay.4.A dynamic timed fuzzy Petri nets(DTFPNs)modelling approach based on the dynamic time delay analysis(e-DTA)is proposed.Firstly,the basic structure of Petri nets is determined by taking advantage of process knowledge.Subsequently,as an improvement,a colored graph describing dynamic time delays between correlated variables is created using data mining techniques.A reachability analysis with temporal constraints is accordingly performed to track the system evolution dynamically.5.A method of industrial process abnormal state monitoring based on DTFPNs is presented.The proposed method is applied to reactor operation optimization,and satisfactory results are obtained,which verifies the feasibility and effectiveness of the proposed method.