Unified Multi-Task Industrial Time-Series Foundation Model
Building an industrial large-model system spanning from foundation model to vertical deployment — a shared backbone plus vertical fine-tuning and prompt engineering — covering monitoring, diagnosis, and prediction tasks in one framework
Project Overview
Modeling monitoring, diagnosis, and prediction tasks separately on the shop floor drives up maintenance cost and hinders knowledge sharing. The laboratory has built a unified multi-task time-series foundation model that integrates industrial process monitoring and transfers to vertical domains via few-shot semi-supervised adaptation.
Research Objectives
A unified multi-task time-series representation backbone
Efficient few-shot fine-tuning for vertical scenarios
Validated deployment for advanced-process quality prediction
Methodology
Time-series signals are aligned with process semantics before entering the large-model backbone; vertical fine-tuning and prompt engineering unify multiple tasks within one model, while a few-shot semi-supervised strategy mitigates the scarcity of industrial labels.
Technical Approach: How It Works
- 1
Time-Series–Semantic Alignment
Sensor time series are windowed and encoded jointly with process semantics (operation, equipment, parameter context), turning time-series segments into an "industrial language" the large model can understand.
- 2
Shared Backbone + Vertical Fine-Tuning
A unified time-series foundation model serves as the backbone; few-shot fine-tuning and prompt engineering adapt it to vertical scenarios such as stamping, die casting, and injection molding, with one backbone serving monitoring, diagnosis, and prediction tasks alike.
- 3
Few-Shot Semi-Supervised Transfer
A semi-supervised strategy exploits the abundant unlabeled data in vertical scenarios, easing industrial label scarcity and enabling cross-scenario knowledge transfer.
Key Results
Deployed at Applied Materials and SAIC-Kaibo among other enterprises, drawing wide media coverage
The direction has produced 4 high-level SCI papers and 3 granted invention patents to date
Intelligent Monitoring of Manufacturing Processes