Multi-attention key-factor-aware convolutional neural network developed for quality prediction of batch processes tackling data sampled at various frequencies

Abstract

Quality prediction is a critical issue in batch processes, where it encounters numerous challenges. Actual batch processes exhibit characteristics of multiple sampling frequencies and multiple stages. The former influences the efficient utilization of data, while the latter typically corresponds to sequential microbial growth stages or operational steps, manifesting as complex process dynamics that affect the effective extraction of process features. This paper presents a multi-attention key-factor-aware convolutional neural network (MKCNN) designed to address both aspects. MKCNN is a multi-branch model, with each branch receiving data sampled at a different frequency as input. Two types of branches are designed: Main Branch and Auxiliary Branch. The former tackles data containing process stage characteristics and local dynamics. In this branch, spatial attention enhances stage-specific features, while channel attention emphasizes the overall local dynamics. The latter handles data covering local dynamics or overall static features. In this branch, either spatial attention enhances local dynamics, or channel attention emphasizes overall static features. Subsequently, features from each branch are fused by a feature decomposition and fusion module (FDFM). FDFM employs cross-attention to capture the correlation among the features from different branches. The proposed MKCNN was evaluated on a real-world ethanol fermentation process (EFP) against support vector regression (SVR), multi-branch convolutional neural network (MCNN), and multi-branch long short-term memory (MLSTM), and so forth. MKCNN demonstrated an average improvement of 11.7% in R² compared to SVR and a 5.7% improvement compared to MLSTM. These results underscore its superior performance in quality prediction.