.png){kind=icon}
.png){kind=icon}
Return
Batch process quality prediction based on TFCMixer modeling
.png){kind=icon}
PRE
AI
DOI：10.1016/j.ces.2026.123901.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Batch processes have dynamic and nonlinear characteristics, and deep learning models suffer from excessively long training times. To address these issues, this paper proposes a batch process quality prediction method based on the Temporal Feature Convolution Mixer (TFCMixer) model. Firstly, the Maximum Information Coefficient (MIC) method is used to screen process variables, retaining only those highly correlated with quality variables. Secondly, to overcome the limitations of traditional time-series feature extraction and significantly shorten training time, a new temporal feature convolution processor is designed. On one hand, it captures temporal dynamic features via lightweight convolutions—accurately tracking data’s temporal evolution while avoiding excessive computation. Meanwhile, dual convolutions handle feature dimensions to reduce redundant parameter interactions, enabling the model to learn inter-feature patterns with fewer iterations, thus analyzing complex feature relationships and shortening the training cycle. On the other hand, a residual scaling block is built into the processor as an adaptive mechanism. By adjusting residual signals’ influence on the final output through a scalable factor, it reduces hyperparameter tuning needs in cross-task scenarios and trial-and-error training rounds, helping the model converge faster on diverse data and further cutting training time. In addition, temporal projection is combined with the processor to achieve seamless time-step transitions, enhancing the model’s ability to handle sequence dependencies. Finally, validation with penicillin fermentation and semiconductor wafer etching data shows the TFCMixer model is effective in prediction accuracy and training efficiency.
Keywords:
TFCMixer
batch process quality prediction
temporal feature extraction
deep learning
training efficiency
Journal
IF:
4.3
Papers:
2.2W
Citations:
5.5W