Generative Convolutional Monitoring Method for Online Flooding Recognition in Packed Towers

Background

Data-driven methods play an important role in monitoring the liquid flooding process for ensuring the efficient and safe operation of packed towers. However, their online recognition performance is often limited due to the imbalanced and nonlinear nature of the flooding data.

Method

In this work, a generative convolutional monitoring (GCM) method is proposed for online flooding recognition. Firstly, a generative model by integrating variational autoencoder with Wasserstein generative adversarial networks is designed to generate information-rich flooding images for enlarging the diversity of the dataset. Secondly, the convolutional neural network is employed for the online recognition of flooding. Finally, feature visualization explains the details of the GCM method in terms of feature extraction. Consequently, the proposed method extracts nonlinear characteristics while overcoming the difficulties associated with unbalanced data.

Significant findings

Experiments on a lab-scale packed tower demonstrate the feasibility of the proposed approach. The flooding state in packed towers can be online detected.