In process industries, time-series forecasting and soft sensing are used to estimate quality variables that are hard to measure online. However, labeled data is scarce, operating regimes change frequently, and retraining models or rebuilding alignment pipelines for each scenario is costly. Existing text-enhanced methods often treat inputs as anonymous numerical columns and fail to leverage the semantic-logical relationships between input variables and prediction targets. To tackle this issue, this article proposes Task-Semantic Field Factorization (TSF), a framework guided by large language models (LLM).
TSF constructs a task-semantic field from task protocols and variable documents before training and utilizes the LLM only for offline semantic construction, while online training and inference remain with conventional time-series backbones. During training and inference, the current numerical window activates variable semantics, allowing semantic information to participate in each prediction and support adaptation to different prediction targets and operating shifts.
On multiple complex industrial forecasting and soft-sensing tasks, TSF reduces MAE by an average of 6.4% in improved settings, with the largest reduction reaching 25.5%. It adds only about 1.8 to 3.0k parameters, with less than 0.008 ms/step of additional online inference overhead. These results indicate that TSF transforms existing process documents into measurable forecasting gains while remaining lightweight for deployment.
Blogger's Review: The TSF framework demonstrates the integration of traditional time-series forecasting with modern language models, enhancing industrial process prediction performance. By incorporating semantic information, TSF significantly improves model adaptability and accuracy, offering new insights for industrial applications. Its lightweight nature also enhances feasibility for real-world deployment.