Approaches To Deep Learning Based Manipulating Strategies Reconstructions For Complex Chemical Processes

At present,a large number of process monitoring and control time series data are collected in production operations of large-scale chemical plants.How to achieve the practical control manipulating experiences and rules is of significance to improve the performance of intelligent automation of plants.The results of previous researches have shown that time series clustering is a very effective way to extract historical manipulating sequences.However,the applicability of the time series clustering method cannot be guaranteed if irregular disturbances happen.In response this problem,the thesis proposes a manipulating strategy reconstruction method(MSR-CNN)using a convolutional neural network model to learn from complex chemical process time series data.The research contents and results obtained in this thesis are presented as follows:1.The influence of multiple disturbance and manipulation variables on key variables in complex chemical processes is investigated.To identify various types of process disturbance states,the advantages of hierarchical agglomerative clustering in dealing with problems that do not require the determination of the number of clusters are analysed,and the Levenshtein distance is used as a time series similarity measure.A symbolic aggregation approximation method is used to process manipulated variable time series,reducing the operational space and computational effort while retaining information on manipulated time series trends.Through numerical examples,it is shown how to pre-process industrial process time series data in an offline manner.2.The practical operating conditions are often different from the historical data and these differences will make the reconstruction of an accurate process strategies more difficult.So this paper uses the hierarchical agglomerative clustering method and a symbolic aggregation approximation method are used to handle historical disturbance variable data and manipulation data in an offline manner,the improved CNN model is also used for deep learning of process disturbance states and the corresponding manipulation sequence strings,and for online reconstruction of the regulatory manipulation strategies.Through the numerical examples and the process industry heat exchanger instance,the adaptiveness and robustness of the proposed deep learning-based method for manipulating sequence reconstruction of complex chemical processes(MSR-CNN)is demonstrated.3.The proposed method is applied to top temperature control of an ethanol water distillation column.The data of key variables,manipulated variables and disturbance variables of the distillation column simulation model outputs are collected,and deep learning and online reconstruction of the manipulation strategies are performed.Assembling control operations under real process disturbance conditions,achieving regulatory control and operation of production process targets.Compared with traditional supervisory control methods,the proposed new method can reconstruct complex chemical process regulation and control strategies well,verifying the effectiveness and superiority of the MSR-CNN method for engineering applications.