|
1 |
Manufacturing Foundation Model (MFM)
The MFM team develops reusable manufacturing foundation models that interpret industrial data and support decision-making across diverse manufacturing systems. We generate and augment data to fill gaps in manufacturing datasets, learn unified representations across signals, images, and text, and adapt the resulting MFM into task- and context-specific MxFMs and intelligent agents.
Manufacturing Data Generation and Augmentation
We generate and augment manufacturing data to capture diverse equipment, operating conditions, and failure modes. This provides broader and more balanced training data when real-world observations are limited.
Data Generation
- MOTIVATION
- High-frequency manufacturing signals are difficult to collect for rare anomalies and Edge/Corner scenarios, while conventional generation methods may not preserve critical high-frequency physical characteristics.
- APPROACH
- We develop a high-frequency signal generation framework for Edge/Corner scenarios that preserves physical characteristics and integrates low-frequency signals and image context through physical filtering.
- IMPACT
- The framework expands data coverage for rare and extreme conditions, enabling robust model development and validation when real high-frequency measurements are limited.
Data Augmentation
- MOTIVATION
- Manufacturing signals vary significantly across machines and operating conditions, while labeled fault data are scarce and conventional augmentation may distort fault-relevant characteristics.
- APPROACH
- We develop manufacturing-specific augmentation using cycle-, order-, and envelope-domain characteristics to preserve fault information while introducing invariance to operating and machine-specific variations.
- IMPACT
- The proposed augmentation improves representation robustness and transferability across machines and operating conditions while reducing dependence on labeled fault data.
Representation Learning and Multimodal Fusion
We learn and fuse representations from diverse manufacturing data to develop a unified MFM. This allows the model to combine complementary information and transfer knowledge across manufacturing systems and tasks.
Representation Learning
- MOTIVATION
- Variations in operating conditions and process stages can obscure meaningful signal patterns, making equipment changes difficult to identify.
- APPROACH
- We learn hierarchical representations that reflect equipment structure and process context, using physical relationships to compensate for operating variations.
- IMPACT
- We aim to support reliable monitoring and diagnosis through representations that preserve meaningful information across changing process conditions.
Multimodal Fusion
- MOTIVATION
- Manufacturing signals and images provide complementary information whose relevance depends on the equipment and task.
- APPROACH
- We combine signals and images through specialized experts, using equipment tags and task context to guide fusion and expert weights to aid interpretation.
- IMPACT
- We aim to improve diagnostic reliability and interpretability by integrating complementary evidence and revealing how different sources contribute.
MFM Adaptation and Manufacturing Agents
From the unified MFM, we develop MxFM variants adapted to specific tasks and operating contexts, along with agents for manufacturing decision-making. This supports flexible deployment in changing manufacturing environments and more informed operational decisions.
MFM Adaptation
- MOTIVATION
- Manufacturing applications require foundation models to adapt with limited labeled data and remain reliable as operating conditions change.
- APPROACH
- We develop adapter-based few-shot adaptation and MLOps workflows for monitoring data shifts, updating models, and validating deployment.
- IMPACT
- Our research aims to reduce adaptation costs and sustain reliable performance through continuous monitoring and model updates.
Manufacturing Agents
- MOTIVATION
- Turning foundation model outputs into manufacturing decisions requires equipment context, engineering knowledge, and substantial expert effort.
- APPROACH
- We develop manufacturing agents that coordinate model outputs, domain knowledge, and analytical tools to recommend evidence-based actions.
- IMPACT
- Our research connects model outputs to shop-floor actions, supporting consistent, traceable decisions in monitoring, troubleshooting, and maintenance planning.
|
