Novel Multi-flow Multi-scale Convolutional Neural Network Developed for Quality Prediction of Batch Processes to Fuse Data With Different Sampling Frequencies

• Published in: International journal of control, automation, and systems Vol. 22; no. 6; pp. 2016 - 2028
• Main Authors: Dong, Yufeng, Yan, Xuefeng
• Format: Journal Article
• Language: English
• Published: Bucheon / Seoul Institute of Control, Robotics and Systems and The Korean Institute of Electrical Engineers 01-06-2024; Springer Nature B.V; 제어·로봇·시스템학회
• Subjects: Artificial neural networks; Batch processes; Batch processing; Control; Data processing; Engineering; Ethanol; Injection molding; Mechatronics; Neural networks; Regular Papers; Robotics; Sampling; 제어계측공학; quality prediction; Attention mechanism; multi-scale feature extraction; batch process; convolutional neural network; multi-branch structure
• ISSN: 1598-6446; 2005-4092
• DOI: 10.1007/s12555-023-0154-8

Quality prediction is a challenging task due to the nonlinearity and complexity of batch processes. In real batch processes, the presence of different sampling frequencies complicates data processing and information digging, making it difficult to fully investigate process information. To address this dilemma, this work developed a multi-flow multi-scale convolutional neural network (MFMSCNN) for the quality prediction of batch processes. MFMSCNN adopts a multi-branch structure to cope with data with different sampling frequencies. A multi-scale feature branch will be adopted to extract the multi-hierarchy features of data containing rich information. Meanwhile, a 1D convolution branch will be applied to the mining process characteristics of data containing less information. Finally, all features in each branch are fed into the fully connected layers to make a quality prediction. In this manner, the process data are fully exploited, and the multi-level features are extracted to better interpret the batch processes. MFMSCNN was evaluated on an industrial ethanol fermentation process and an injection molding process. It obtained remarkable performance on both batch processes. The prediction results of the proposed method are superior to many other methods.