Dynamic Representation Learning And Changing Operating Condition Monitoring With Dynamic And Static Coordination For Industrial Processes

With the development of information technology,the industrial process has entered a new era of rapid development based on the concept of intelligent manufacturing.The continuous expansion of industrial production scale and the significant improvement of automation level put forward higher requirements for the safety,stability and environmental protection of processes.The use of effective process monitoring methods to identify process faults in time is crucial to ensure the safe and efficient operation of industrial processes.In recent years,data-driven process monitoring methods have developed rapidly and have received extensive attention.However,the dynamic characteristics of modern industrial processes and changing operating conditions,including normal switching of operating conditions and shortterm fluctuations,have brought great challenges to the existing research work.Based on this,this paper studies the dynamic characteristics of complex industrial processes and changing operating condition identifications method with dynamic and static coordination,aiming to accurately identify the complex and diverse operating conditions of the process and the switching between different states.This paper contains five parts.First,typical process monitoring methods based on dynamic representation learning are reviewed,and their problems and challenges are briefly summarized.Then,the problem of representation of process dynamic characteristics and identification of changing operating conditions under the influence of control action is solved.On this basis,the fourth chapter realizes the coordinated identification of the state changes of process variables and performance indicators.Furthermore,the problems of changing operating condition monitoring in two complex scenarios with irregular sampling interval and missing variable values and large-scale non-stationary are considered.The specific research content of the article is as follows:To deeply understand the existing dynamic representation learning methods,the second chapter systematically analyzes and compares six typical process monitoring methods based on dynamic representation learning.First,the objective functions,modeling procedures and fault detection strategies of typical methods are reviewed.Then,from the perspective of theoretical analysis,the properties of each method are summarized and compared,including the construction of data matrix,the statistical properties of latent variables,and whether the separation of dynamic and static information is realized.Finally,the three-phase flow process data are used to compare and verify the performance of each method,including model generalization performance,fault detection time,fault detection rate,etc.Through the dual analysis of theory and application,the dynamic characterization ability and fault detection performance of each method are revealed.Aiming at the difficulty in identifying complex state changes such as transient fluctuations in the process and switching of operating conditions under the regulation of control action,Chapter 3 proposes a dynamic and static cooperative changing operating conditions monitoring method based on an enhanced canonical variable analysis.First,the enhanced canonical variable analysis algorithm is designed to extract latent variables with strong dynamic correlation and slow change from process data.The algorithm can effectively separate dynamic and static information while extracting effective dynamic representation.Then,a dynamic and static coordinated changing operating condition identification strategy is proposed,and statistical indicators are established to separately monitor the process dynamic and static information and the process change speed.Different from the traditional method that only considers open-loop data,the proposed method can comprehensively and clearly indicate the influence of the control action on the process,and accurately distinguish the normal process switching and real faults of the process.Experiments in thermal power generating units show that the fault detection time of the proposed method is 79 sampling points earlier than other methods,and the fault detection performance is significantly improved.Aiming at the problem that the key performance indicators of process are difficult to measure and the corresponding complex state changes are difficult to identify,the fourth chapter proposes a performance-relevant full decomposition of slow feature analysis and bidirectional identification strategy for process space decomposition and condition monitoring.First,slow feature analysis is used to extract key dynamic representations from the process data.Then,by analyzing the dynamic relationship between key performance indicators and process data,the process data is decomposed into performance-relevant and process-relevant subspaces.After that,the information of each subspace is mined by the bidirectional identification strategy.On the one hand,the two subspaces are monitored separately such that the changes of the process states and key performance indicators can be simultaneously monitored.On the other hand,the dynamic and static information of each subspace is monitored collaboratively,then the key performance indicators and process state changes under the influence of closed-loop control are comprehensively depicted.The experimental cases show that,the proposed method timely indicate abnormal changes of performance indicators and accurately identify the normal switching of performance indicators set and process operating conditions.To address irregular sampling intervals and missing values,the fifth chapter proposes an interval-aware probabilistic slow feature analysis algorithm.Different from the traditional method with a fixed mathematical model,the proposed method introduces a time interval function to flexibly adjust the influence of the historical data on the current process state according to the time intervals.At the same time,different interval functions are designed to accommodate the changing law of the dynamic relationship between adjacent samples in different scenarios.Then,the expectation-maximization algorithm combined with the interval-aware Kalman filtering is deduced to estimate the parameters of the proposed algorithm while overcoming the information loss brought by missing values.Based on the estimated state variables,process error and varying speed,three statistical indicators are constructed to comprehensively identify changes in process operating states.It is the first time that the interval-aware idea is involved into the state-space model.And the law of process dynamics changing with sampling intervals is deeply explored.The experimental results in the Tennessee process show that the average false alarm rate of the proposed algorithm is 0.67%,the average fault detection lag is 6 sample points,and the average fault detection rate is 86.40%.To solve the problem that the operating conditions of industrial processes are frequently switched driven by multiple factors and exhibit significant non-stationary transient characteristics,the sixth chapter proposes a coarse-to-fine condition identification method.In the coarse-level decomposition,the condition indicators are found,and the density-based clustering algorithm is used to explore the coupling relationship between the indicators from both time and spatial dimensions.In this way,the process data can be divided into different condition clusters.Next,an in-depth analysis of the impact of indicator variable on process dynamics is explored to decompose condition clusters into fine-grained condition modes.The non-stationary characteristics within each condition mode are reduced such that local models are established to comprehensively identify the fluctuation of process dynamic and static information under different modes.The proposed method explores the changing law of the process dynamic characteristics influenced by the mutually coupling indicator variables,and provides a clear physical explanation for the identified condition modes.Meanwhile,the identification of condition mode division can be automatically realized.In the case studies,the condition modes identified by the proposed method are consistent with the real situation,and the average fault detection lag is shortened to 100 sampling points.