Deep Learning And Its Application To Flooding Prediction In Packed Towers

Packed columns have been widely used in chemical separation processes,including absorption and distillation.As known,gas and liquid flow rates in packed towers are limited by the tendency of the columns to flood.The flooding phenomenon often leads to poor column efficiency,and may even shut down the entire production line.On the other hand,high gas velocity often means high column capacity.With the demand of efficiency,the column is desirable to be operated close to the flooding point.Therefore,in order to ensure an efficient and a safe operation of packed columns,research on online monitoring the occurrence of flooding becomes necessary in practice.In this thesis,the traditional methods of flooding monitoring in packed towers are first reviewed.As a novel nonlinear modeling method,the deep learning is applied to extract important features and analyze the flooding phenomenon.To better capture different process characteristics,information in process variables and images are integrated and introduced into the deep learning model.Additionally,detailed implemented steps of online flooding monitoring are proposed.Consequently,by identifying different states of a packed column,online flooding monitoring can be achieved.The main contributions of this work are as follows.(1)Traditional prediction models for flooding monitoring only extract linear features.The deep belief networks(DBN)is applied to predict the flooding in packed towers.Additionally,an adaptive DBN(ADBN)is proposed to select its parameters in an efficient manner.Consequently,the flooding prediction model can extract nonlinear features more quickly.The experimental results show that the ADBN model is superior to the traditional models.(2)To analyze the operational state of packed tower more deeply,using both of the information in color images and process variables,a flooding prediction model is developed.With the nonlinear characteristics extracted in informative data,a convolutional neural network(CNN)is proposed to judge the operational state in a packed tower directly.Consequently,the integrated model shows more accurate and reliable performance in online identifying the occurrence of flooding.