Identification And Deduction Of Abnormal Working Conditons In FCC Fractionation System Based On Data And Knowledge Fusion

Fluid catalytic cracking(FCC)is an important unit for petrochemical industry.Abnormal working conditions in its production process may affect product quality in minor cases or cause safety accidents in serious cases.The identification and deduction of abnormal working conditions in FCC process can help to improve the safety and stability of the system and avoid major accidents.A data driven and knowledge fusion method is proposed to identify and deduct abnormal working conditions in FCC fractionation system.With the key variables of the system as the core,the abnormal working conditions of the unit are monitored,identified and deducted.The main contents of the method are as follows:(1)A anomaly identification method based on the prediction of the explicit variable is proposed to early identify the abnormal working conditions of the system by using the key explicit variables as monitoring points.Firstly,the complex network relationship between system variables is constructed and the centrality of each node is calculated.The key variables(i.e.,the target variables)of the FCC fractionation system of the anomaly identification model are identified.Secondly,the correlation coefficient is used to screen the feature variables to reduce the input dimension and remove the redundant variables.Finally,the attention mechanism,convolutional neural network and long short-term memory network are organically fused to train and form an anomaly identification model for the explicit variable.(2)An abnormal working condition identification method based on soft sensing of implicit variables is proposed to identify abnormal working conditions of fractionation tower and heat exchanger equipment based on real-time soft sensing of implicit variables.Firstly,the abnormal frequency of each part of the equipment is analyzed based on the alarm value statistics of each process point and the operation mechanism of the plant,and the key implicit variables are identified.Secondly,the typical working conditions of the fractionation system are focused using the historical data as the initial conditions for the mechanism modeling.Thirdly,the mechanism simulation model of the fractionation system is established to obtain the correlation data sets of the implicit variables and the process process variables.Finally,the process features of the fractionation system are extracted based on the deep neural network structure built in(1).The online soft sensing model of the implicit variables is obtained by training the neural network with the correlated data set to identify equipment anomalies in time.(3)A knowledge graph-based method for the deduction of abnormal operating conditions is proposed.Based on the characteristics of FCC fractionation unit,a knowledge graph model of the system is established based on cause-effect analysis,which is used to show the influence relationship among variables when abnormal working conditions occur in the fractionation system.In this paper,the output value of the anomaly identification model is used to identify the abnormal behavior of key variables.The abnormal deduction of the fractionation system is carried out through the established knowledge graph to find the root cause of the anomaly and the abnormal working conditions that may occur.The knowledge graph established provides guidance for anomaly disposal.Through deep integration of mechanism knowledge,process simulation,big data and deep learning technologies,this paper studies variable screening,parameter modeling,anomaly identification and deduction tracing.This work forms an integrated solution from localization,monitoring,identification to deduction for the abnormal working conditions of FCC fractionation system.The proposed neural network integration framework applied to explicit variable anomaly identification can identify the abnormal behavior in advance with root mean square error less than 5%.The constructed tower plate efficiency and heat exchanger thermal resistance models have root mean square error of 0.859% and 0.19%,respectively,which can achieve real-time anomaly identification of implicit variables.T he constructed knowledge graph model contains all key variables within the process and can accurately evaluate the anomaly propagation path and locate the fault source.