Complex Nonstationary Industrial Processes Modeling And Monitoring Based On Data-driven Methods

Process monitoring is an important technology to ensure the safe operation of industrial processes and ensure product quality.With the rapid development of sensor technology and computer technology,process data has been easily obtained.Therefore,data-driven process monitoring methods have developed rapidly.The multivariate statistical process monitoring(MSPM)methods have advantages of simple model and easy implementation when dealing with high-dimensional data.MSPM methods have attracted attention of academia and industry.Although the multivariate statistical methods have been rapidly developed in the field of process monitoring and have achieved great research results,the existing work is based on the assumption that the processes are stationary.However,many process variables are nonstationary whose statistical characteristics are changing with time due to many factors such as market demand,throughput changes,unmeasured disturbances,etc.The industrial processes show the nonstationary characteristics.Fault detection and diagnosis for nonstationary processes is a difficult task since the fault signal may be buried in the nonstationary trends of variables.This thesis deeply studied the characteristics of nonstationary processes based on the previous.In order to solve the practical problems,a series of algorithms for statistical modeling,online fault detection,and online fault diagnosis for nonstationary processes are proposed,which are summarized as follows:(1)For the incipient fault detection in stationary and nonstationary hybrid processes,a hierarchical modeling strategy is proposed.In the lower level,the cointegration analysis is used to extract the long-term equilibrium relationship among nonstationary variables,and the nonstationary variables monitoring model is constructed to detect the incipient fault covered by the nonstationary trends of variables.For the stationary variables,the principal component analysis is used to extract the characteristic information in the stationary variables,and an incipient fault detection model is established to detect the incipient fault covered by the noise in the stationary variables.In order to evaluate whether the correlation between stationary and nonstationary variables is affected by incipient fault,the characteristics of the lower nonstationary variables and stationary variables are combined to construct the upper level monitoring model.Two-level of three monitoring models constitute the incipient fault detection model,which can timely detect the incipient fault.(2)To handle the monitoring problem in the large-scale nonstationary processes with large number of process variables,an automatic variable sub-blocks decomposition method and a distributed modeling strategy for refined monitoring are proposed.The nonstationary process variables are assigned into the different blocks through an iterative variable selection algorithm based on sparse cointegration.The local nonstationary characteristics can be well encapsulated in each block.The distributed monitoring strategy is proposed to reflect the local behaviors of the process by describing the long-run equilibrium relationships among the nonstationary variables in each block.Finally,a new information fusion method is proposed to combine the feature information of different blocks,and a global model is constructed to evaluate the correlation among different blocks.It is the first time that realize the distributed modeling and refined monitoring for the large-scale nonstationary processes based the above methods.(3)The closed-loop feedback system's regulation effect on the process and its impact on process monitoring are discussed in detail to solve the fault detection problem of closed-loop system in nonstationary processes.A concurrent modeling and monitoring method is proposed based on dynamic and static feature decomposition.It is the first time that cointegration analysis enables separate description of the static and dynamic equilibrium relationships for distributed monitoring the large-scale nonstationary processes under closed-loop control.The impact of closed-loop feedback control on process can be eliminated by the concurrent modeling,which can distinguish the operation changes and real fault.(4)To solve the problem of online fault isolation in nonstationary processes,a sparse reconstruction strategy is proposed.The fault variable isolation problem is transformed into an automatic variable selection problem by integration LASSO operators into the fault reconstruction model.The proposed method can automatically and real-time isolate multiple faulty variables that are responsible for the abnormal operation for the nonstationary processes without using any historical fault data.The above methods provide new ideas and solutions for various problems encountered in the complex nonstationary industrial processes.The proposed methods are applied to the nonstationary processes such as 1000-MW ultra-supercritical thermal power unit and three-phase flow facility and are compared with several monitoring methods to demonstrate the effectiveness of these new approaches.Finally,some future research studies are discussed based on the conclusion of the thesis.