关键词: 间歇过程, 质量预测, 奇异谱分析法, 时域卷积网络, 多头自注意力机制

Abstract: To improve the stability of Time-domain Convolutional Network (TCN) training when batch size changes, and to address the issue of batch process quality prediction failing to capture long-term dependence and global correlation, which led to poor prediction accuracy, a model for batch process quality prediction was proposed. Batch Group Normalization (BGN) and Mish activation function were incorporated to enhance the residual structure of the time-domain convolutional network with a multi-head self-attention (BMTCN-MHSA) mechanism. Firstly, the three-dimensional data of the batch process was unfolded into a two-dimensional matrix form and normalized. Singular Spectrum Analysis (SSA) was then introduced to decompose the reconstruction. BGN was added to the residual part of the time-domain convolution to reduce the network model's sensitivity to batch size changes. The Mish activation function was introduced to improve the model's generalization ability, and the multiple self-attention mechanism was utilized to correlate and assign weights to feature information at different sequence locations, further extracting key feature information and interdependence relationships in the sequence to better capture the dynamic characteristics of the batch process. Finally, the experimental validation is carried out using penicillin simulation experimental data, and the experimental results show that compared with the time-domain convolutional network, the mean absolute error of the BMTCN-MHSA model is reduced by 56.86%, and the mean square error is reduced by 48.80%, and the r-squared reaches 99.48%, which suggests that the model proposed in this paper improves the accuracy of the quality prediction.

Key words: batch process, quality prediction, singular spectrum analysis, time-domain convolutional network, multi-head self-attention mechanism