Data-Driven Modeling And Monitoring Of Transient Phases In Continuous Industrial Processes

Due to variations in set-points,changes in material,seasoning and the aging of equipment,the condition of processes changes frequently with different degrees.Therefore,these changes should be identified for the purpose of multimode process modeling and fault detection.In most researches of multimode process monitoring,stable modes were often emphasized while transitions between stable modes were always neglected.Industrial transitions usually contain intense fluctuations,which might result in serious systematic faults.Therefore,the fault detection for transitions should be considered as an important issue in process monitoring and controlling.In this paper,the transient processes are discussed from different perspectives according to different statistical characteristics including non-linearity,auto-correlation and non-Gaussianity.The main content of this paper can be summarized as follows:1)Consider the non-linearity in transitions.Commonly,non-linearity is considered as one of the basic characteristics in industrial processes.In order to handle the non-linearity in the processes,we proposed a Distributed Model Projection(DMP)based multimode identification and fault detection algorithm.According to different residuals with respect to specific models,a multimode process can be clustered into several clusters.Every stable mode is considered as one cluster and can be described using one stable model.While a transition is separated into several sub-segments,each of which can be characterized using one stable model.On the basis of offline identification results,the online identification and fault detection are carried out then.The online model is updated so that the online identification and fault detection can be implemented simultaneously.2)Consider the non-Gaussian characteristics in transitions.Because of the sophisticated features of a transition,the process data can hardly be Gaussian distributed.Inspired by the concept of Just in Time Learning(JITL),we proposed a fault detection method for transitions with non-Gaussian distribution.Firstly,the difference between independent components of different data is measured by mutual information(MI)method.Then,in the online fault detection step,the training data is updated by the ones which have close mutual information with the monitored data blocks.The control limits are calculated based on the training data by using the kernel density estimator.After that,the difference between the monitored data blocks and their corresponding training data is measured by the different independent components between them.Finally,the fault detection work is executed by comparing the corresponding difference between the statistics and the control limit.3)Consider the dynamic characteristics in transitions.The system condition changes from one stable mode to the other one.The variations of system characteristics are therefore directional,which results in auto-correlations in the transient process.We propose a Dynamic Mutual Information Similarity(DMIS)based multimode identification and fault detection algorithm.According to the dynamic information within each part of a multimode process,the whole process is divided into several sub-segments,in which dynamic characteristics remain stable.Considering the difference between stable modes and transitions,two kinds of moving window strategies are proposed.Online identification and fault detection method are then carried out using a series of DPLS models.The fault could be detected immediately using the updated models.