Fault detection and diagnosis for reactive distillation based on convolutional neural network. (February 2021)

• Record Type: Journal Article
• Authors: Ge, Xiaolong; Wang, Beibei; Yang, Xinchuang; Pan, Yu; Liu, Botan; Liu, Botong
• Is Part Of: Computers & chemical engineering. Volume 145(2021)
• Publication Date: 2021-02
• Subjects: Reactive distillation -- Process intensification -- Fault detection and diagnosis -- Control -- Convolutional neural network; Chemical engineering -- Data processing -- Periodicals; 660.0285
• DOI: 10.1016/j.compchemeng.2020.107172
• Languages: English
• ISSNs: 0098-1354

Abstract

Highlights: Process intensification for formic acid production was realized through reactive distillation. Genetic algorithm combined with user-defined kinetics model effectively tackle the optimization difficulty. Features in both spatial and temporal domain under PI control are taken into account with measurement noise to formulate samples. Convolutional neural network was employed to diagnosis the operating state of reactive distillation. Machine learning information in each layer was visualized using t-SNE. Abstract: Reactive distillation (RD) shows its strength in achieving process intensification. However, the complex phenomena integrated in RD usually leads to various abnormal operating states, e.g. catalyst deactivation. Although control schemes have been designed to tackle some disturbances, diagnosing the operating state online is of vital importance for effectively avoiding serious accidents. In the present work, by using intensified process for formic acid production as benchmark, optimal design with stochastic algorithm was firstly performed and dynamic test was carried out to validate effectiveness of control structure. Then thirteen practical faults were considered and the corresponding response was simulated. By considering features in both spatial and temporal domain, historical dynamic process data with measurement noise was used to formulate samples, based on which deep convolutional neural network was trained and validated. The machine learning information in each layer was visualized using t-SNE and fault diagnosis rate shows the significance of the method. Graphical abstract: Image, graphical abstract