CAREER: Transforming Machine Learning Models Developed in Labs to Manufacturing Plants for In-Process Quality Prediction

自动化、数据科学和人工智能等领域的最新进展正在为先进制造业创造新的机会。然而，大多数基于机器学习的解决方案都是在实验室环境中开发的，这需要在制造工厂中实施大量的模型调优和昂贵的数据标记。技术障碍来自实验室和工厂数据在数量、分布、准确性和模式上的主要差异。该学院早期职业发展（Career）奖将研究新的机器学习方法，使机器学习普遍化和可部署。如果成功，该项目将加速人工智能在制造工厂的部署，并降低中小型制造商进入工业4.0的门槛。该项目还有望通过吸引初高中学生和当地产业，为新的制造业劳动力的发展做出贡献。该项目旨在开发具有可扩展模块的机器学习架构，以从大量未标记数据流中学习，并适应工厂中动态变化的制造条件。实验室到植物的转换将在验证两个科学假设的基础上实现：(1)用于表征大量未标记数据的通用模型可以有效地学习植物数据的相似性；(2)已建立的模型可以完全适应不可见但相关的场景，只需有限的调整。将配置一个基于变压器架构的新型机器学习框架，以同时实现：i)从大量工厂数据中进行任务不可知的自监督对比学习，以实现多层次数据表征；Ii)标准化流程，用于在传感数据之间建立一对一的映射，从而在工厂中生成虚拟传感数据，以改进质量预测；iii)在不同的制造条件下有效和高效地适应模型，提示模型翻转。如果成功，该项目将实现通用的、可部署的机器学习解决方案，这些解决方案将很容易扩展到广泛的制造应用中，并帮助美国制造商加速采用智能制造技术。该项目由先进制造计划和促进竞争研究的既定计划（EPSCoR）共同资助。该奖项反映了美国国家科学基金会的法定使命，并通过使用基金会的知识价值和更广泛的影响审查标准进行评估，被认为值得支持。

Recent advances in areas such as automation, data science, and artificial intelligence, are creating new opportunities for advanced manufacturing. However, most machine learning-based solutions are developed in lab environments, which require extensive model tuning and expensive data labeling to be implemented in manufacturing plants. The technical barrier arises from the major discrepancies in the amount, distributions, veracity, and modality between lab and plant data. This Faculty Early Career Development (CAREER) award will investigate new machine learning methodology to make machine learning generalizable and deployable. If successful, the project will accelerate the deployment of artificial intelligence in manufacturing plants and lower the entrance barrier to Industry 4.0 for small and medium manufacturers. This project is also expected to contribute to the development of new manufacturing workforce by engaging middle/high school students and local industries. This project aims to develop a machine learning architecture with expandable modules to learn from massive unlabeled data streaming and adapt to dynamically changing manufacturing conditions in plants. The lab-to-plant transformation will be realized upon testing two scientific hypotheses: (1) a generic model for characterizing massive unlabeled data can effectively learn the similarities of plant data; (2) an established model can be fully adapted to unseen but related scenarios with limited tuning. A transformer architecture-based novel machine learning framework will be configured to simultaneously realize: i) task-agnostic self-supervised contrastive learning from massive plant data for multi-level data characterization; ii) normalizing flow for building one-to-one mapping between sensing data toward virtual sensing data generation in plants for improved quality prediction; and iii) prompt model turning for effectively and efficiently adapting models between different manufacturing conditions. If successful, this project will enable generalizable, deployment-ready machine learning solutions that will be readily scalable for a broad scope of manufacturing applications and help U.S. manufacturers adopt smart manufacturing technologies at an accelerated pace.This project is jointly funded by the Advanced Manufacturing Program and the Established Program to Stimulate Competitive Research (EPSCoR).This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.