Distributed Fault Diagnosis Methods Of Complex Industrial Process

With the rapid development of modern technology and big data technology,industries such as machinery,energy,petrochemical,transportation,and defense that incorporate high-tech and intelligent equipment are becoming increasingly large-scale,integrated,high-speed,automated,and even intelligent.What follows is the ever-increasing requirements for the safety and reliability of industrial process operations.The process industry under the background of large-scale and big data has resulted in a large number of on-site operation units and a huge amount of collected data.Only relying on regular maintenance detection and data alarm analysis is not only resource-consuming but also inefficient.Traditional manpower control and abnormal determination are far from being able to meet the monitoring needs of modern complex industrial processes.At the same time,there are many intricate,interrelated and coupled interrelationships between and within the components of the industrial system.When a system fails,one type of failure may have multiple symptoms,and one type of symptom is often caused by multiple failures.It is difficult to accurately diagnose the failure only by relying on a single theoretical method and information source.Therefore,it is necessary to combine the mechanism of industrial process and historical big data to study fault detection and diagnosis technology to realize the condition-based maintenance of the system.If a fault in a complex engineering system is detected and even traced back to the source of the fault,and the location or scope of the fault is determined,the optimal maintenance strategy can be formulated based on various factors such as current personnel and cost,thereby ensuring the safety of the complex engineering system.This paper starts from historical industrial big data,integrates process mechanism knowledge,and builds an overall architecture of distributed process monitoring and fault fusion diagnosis.It solves and breaks through the relevant basic theories and key technologies in the field of data-driven process monitoring and fault diagnosis under the distributed architecture.It focuses on theoretical research and application verification of the health status assessment and fault traceability of complex engineering systems.The main content of this paper includes the following five aspects:(1)For the alarm prediction in the independent subsystem,a centralized process monitoring method based on the combination of multivariate causal analysis and Bayesian network parameter learning is proposed.The incomplete mechanism information of the system is combined with the historical data generated by the operation of the system to establish a Bayesian network model for alarm prediction and analysis.Based on the historical data of the process,the alarm events are analyzed and reasoned,and the prediction of single-variable alarm and multi-variable alarm events is realized.(2)For fault monitoring and source tracing in subsystems,a graphical model establishment method combining causality analysis and non-parametric probability density estimation is proposed.The proposed method overcomes the obstacles of traditional Bayesian network processing time series data.It has stronger applicability,because it does not require discretization processing or distribution function assumptions under the traditional Bayesian network method.The proposed multivariate causal analysis method is used to establish the causal structure of the system,and the kernel density estimation method of non-parametric estimation is used to establish the probability connection relationship between variables.In order to ensure the accuracy of probability density estimation,an index for evaluating the accuracy of probability density estimation is proposed.Through reasonable optimization of parameters,the probability estimation under optimal conditions is obtained.The abnormal events of the system can be detected by the change of the probability density function between the variables,and the root cause of the fault can be traced and located by Bayesian inverse inference.(3)Based on the established independent subsystem modeling and fault diagnosis strategy,a multi-layer distributed system structure model is established.It realizes the communication of distributed objects and solves the problem of fault detection and diagnosis of large-scale systems.A plant-level distributed monitoring three-tier architecture for complex process industries is proposed:dimensionality reduction and decomposition of complex data,partial process monitoring of subsystems and global integrated diagnosis of the entire process.In the process of strategy implementation,the collected process data is integrated with process mechanism knowledge,model,structure and related knowledge.A process optimization decomposition strategy based on detection performance is proposed to decompose the system into several interconnected subsystems on the basis of ensuring the detection performance,a causal probability graph model is established for each subsystem.Through the shared variables between the subsystems,the operating units are connected into a complete large-scale system,which realizes the theory and simulation experiment verification of related data processing,fault detection,causal reasoning and fault diagnosis.(4)Combined with the actual industrial situation,the system is optimized and decomposed to obtain the set of shared variables that connect the subsystems.These variables are the bridge of information transmission between systems.Considering the physical connection and information transmission between subsystems,on the proposed distributed three-tier architecture,a distributed partial least squares algorithm is proposed to solve the problem of information interaction between subsystems and complete the monitoring of subsystems.Next,the information design Bayesian fusion index of all subsystems is integrated in the fusion center to make a global decision diagnosis.In addition,its effectiveness is verified by experiments and comparisons with a variety of existing related solutions.(5)Combining model-driven and data-driven methods to solve the obstacles to the coupling of system states in the model.The accuracy of modeling and process monitoring is guaranteed while the information exchanges between subsystems.The subspace identification and residual design methods of large-scale interconnected systems are proposed for fault detection.According to the characteristics of the connection between subsystems,the unknown state coupling of the system is transformed into the known observation coupling.The subspace identification technology and LQ decomposition solution strategy are used to obtain the system model,and then the fault detection residual generator and threshold are designed to solve the system identification and fault detection problems in the case of system matrix coupling.