Deep Learning And Its Application To Flooding Monitoring In Packed Towers

As an important gas-liquid mass transfer equipment,packed towers have been widely used in industrial productions.Generally,plant safety and efficiency require reliable process control in the distillation and absorption applications.Nevertheless,the occurrence of flooding decreases efficiency when the vapor flow disrupts the liquid flow.When unchecked,flooding may disrupt the entire production process and even lead to product losses and system downtime.Therefore,it is promising to develop efficient flooding monitoring methods using the recent deep learning methods.Traditional flooding monitoring methods in packed towers are reviewed in this thesis.As a novel modeling method,the deep learning is applied to establish data-driven models for flooding monitoring.First,an improved model,i.e.,three dimensional long short term memory neural network(3D LSTM NN)based on recurrent neural network is proposed,which can extract time features of process variable data.However,only monitoring the differential pressure is still difficult to identify flooding.To overcome the problem,a convolutional long short term memory neural network(ConvLSTM NN)is proposed to extract spatiotemporal features of video data in packed towers.Finally,a deep learning flooding monitoring method is proposed.The main contributions are as follows:(1)The differential pressure is a key factor that indicates flooding phenomenon in packed towers.Only using a data-driven recurrent neural network model to monitor the differential pressure is still difficult to achieve satisfied results.To improve the monitoring performance,an improved model,i.e.,3D LSTM NN is proposed,which can extract time features of process variable data.The experimental results show that 3D LSTM NN can obtain better and more reliable prediction performance.(2)A novel monitoring method for flooding supervision based on the video data of packed towers is proposed.First,a convolutional neural network is formulated,which can extract spatioal features of image data.Moreover,combining a long short term memory,a ConvLSTM NN is proposed,which can extract spatiotemporal features of video data.Consequently,a practical flooding monitoring method is developed.The experimental results show that the method can identify flooding effectively.