Data Quality in Manufacturing Industrial Internet Integration

该研究项目的重点是数据质量评估和改进，以便在制造业中有效地部署人工智能（AI）。在现代制造业中，人工智能指导的决策已经彻底改变了生产，产品质量，设计定制和制造可持续性。然而，制造人工智能模型缺乏定量数据质量评估和有效的数据准备方法，这带来了一些关键挑战，包括不可信的人工智能决策，处理大型但质量差的数据集的高能耗，以及缺乏更有效的数据集在制造系统之间共享以进行人工智能模型训练。这些挑战大大减缓了人工智能技术在制造业中的应用，从而严重影响了美国制造业的全球竞争力。 该研究项目定义和评估定量制造数据质量指标，基于制造特征推进数据质量保证的科学知识，并促进人工智能建模的数据集准备。 因此，该研究不仅可以通过提高制造数据质量来快速训练和比较人工智能模型，还可以减少环境对数据计算、通信和存储的影响。该研究项目还包括针对大学生和制造业劳动力发展的全面推广和教育计划，包括小组讨论，面向代表性不足的学生和从业人员的推广研讨会以及制造业人工智能竞赛。 该研究项目的目标是定义，评估和提高数据质量，以使异构机器，传感器和计算设备集成的制造工业互联网中的数据集兼容使用。该项目建立了数据质量方法，以解决基于不同制造布局的制造特定数据格式和模式的挑战。首先，数据质量被定义为与AI模型性能的方差成反比。潜在神经推荐系统研究数据集和AI模型之间的接口，以评估使用不同AI模型时的数据质量。第二，制造数据质量建模的基础上，独特的制造数据特征，从不同的制造布局和数据形态的图形。第三，在确定数据质量差的根本原因后，通过上下文强盗的集成主动学习生成黄金数据集，以确保制造AI模型对数据源可变性的鲁棒性能。数据质量方法与制造层次变量关系、多模态数据和具有有效特征表示的布局表示相关联。该奖项反映了NSF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。

This research project is focused on data quality evaluation and improvement for effective Artificial Intelligence (AI) deployment in manufacturing. In modern manufacturing industries, AI-guided decision-making has revolutionized production, product quality, design customization, and manufacturing sustainability. However, the lack of quantitative data quality evaluation and effective data preparation methodologies for manufacturing AI models pose a few critical challenges, including untrustworthy AI decision-making, high energy consumptions to process large but quality-poor datasets, and a lack of more effective datasets to be shared across manufacturing systems for AI model training. These challenges greatly slow down the adoption of AI technologies in manufacturing industries, thus significantly impacting the global competitiveness of US manufacturing. This research project defines and evaluates quantitative manufacturing data quality metrics, advances scientific knowledge for data quality assurance based on manufacturing features, and promotes dataset preparation for AI modeling. As a result, the research not only enables fast training and comparison of AI models due to improved manufacturing data quality, but also reduces environmental impact on data computation, communication, and storage. This research project also includes a comprehensive outreach and education program for college students and manufacturing workforce development, including panel discussion, outreach seminars to underrepresented students and practitioners, and manufacturing AI competitions. The goal of this research project is to define, evaluate, and improve data quality to enable compatible usage of datasets in Manufacturing Industrial Internet integrated by heterogenous machines, sensors, and computation devices. The project builds the data quality methodology to address the challenges based on manufacturing specific data format and modalities from different manufacturing layouts. First, the data quality is defined as inversely proportional to the variance of AI model performance. A latent neural recommender system investigates the interface between datasets and AI models to assess data quality when different AI models are used. Second, manufacturing data quality is modeled based on the unique manufacturing data features from graphs of different manufacturing layouts and data modalities. Third, after the root causes of poor data quality are identified, golden datasets are generated by ensemble active learning by contextual bandits to ensure robust manufacturing AI model performance to data source variabilities. The data quality methodology connects to the manufacturing hierarchical variable relationship, multimodal data, and layout representations with effective feature representations. Methodologies are validated by both real datasets in Semiconductor Manufacturing and a Manufacturing Industrial Internet testbed.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.