Large Language Model-driven Two-Stage Spatiotemporal Dual-Stream Collaborative Framework for Industrial Quality Prediction
Year: 2026
Type: conference-paper
Abstract: Addressing the challenge of indirectly and accurately measuring hard-to-predict key process variable in complex industrial processes, data-driven prediction methods have emerged as a key solution. However, existing approaches face limitations in complex operating conditions, including insufficient few-shot modeling capabilities, and weak representational understanding. In recent years, Large Language Models (LLMs) have demonstrated significant potential due to their strengths in cross-modal knowledge transfer and few-shot learning. Inspired by theses, this paper proposes a LLM-based two-stage spatiotemporal dual-stream collaborative key process variable prediction framework (LLM-TS2C). Specifically, LLM-TS2C comprises a Temporal Patch Dynamic-Aware Branch (TPDB) and a Spatial Graph Dependency-Aware Branch (SGDB), which jointly model long-range temporal dependencies and multivariate cross-channel couplings via a two-stage training framework, including a representation learning pre-training stage and a key process variable prediction stage. To mitigate representational discrepancies across modalities and enhance spatiotemporal interactions, LLM-TS2C introduces channel-wise and patch-wise differentiated tokenization strategy. Moreover, a cross-branch feature interaction strategy is proposed to enable spatiotemporal collaborative modeling. Finally, experiments conducted on the air preheater dataset further validate the superiority of the proposed method, with significant improvements observed in the models representation and prediction capabilities under both full-data and few-shot scenarios.
Authors: Zehao Fu, Han Liu, Shuo Tong, Runyuan Guo
Institution: Xi'an University of Technology
Language: English
Subfield: Industrial and Manufacturing Engineering
Field: Engineering
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7170173673