EAGER: CoFedAI: Cost-sensitive Federated AI for Smart Manufacturing Data-Sharing

Virginia Polytechnic Institute And State University, Blacksburg VA

Investigators

Abstract

This EArly-concept Grant for Exploratory Research (EAGER) project will investigate a cost-sensitive data-sharing paradigm that integrates manufacturing data from multiple manufacturers to improve supervised learning in smart manufacturing. This effort addresses a critical problem in the implementation of Artificial Intelligence (AI) in US manufacturing, since AI methods benefit from training on large datasets and manufacturers typically keep their data secret. The investigators will research methods for preserving the privacy of that data, with the goal of enabling a future manufacturing service infrastructure to aggregate, manage and reuse data from multiple manufacturers. Such an infrastructure can benefit manufacturing by establishing a data-sharing marketplace that enables domestic partnerships and accelerates the adoption of AI technologies, thus enhancing the international market share of the United States. Aspects of this work will also be incorporated into the courses taught by the PIs. The project lays the foundation for a manufacturing data-sharing ecosystem by creating task-specific similarity metrics and a methodology to differentiate the contributions from multiple manufacturing data owners. The selection of suitable data sources for data aggregation depends on the categorization of the data derived from the various sources for similarity in data distribution and variable relationship. The Cost-sensitive Federated AI (CoFedAI) framework will facilitate data exchange to unlock the value of knowledge transfer for AI in manufacturing by employing a cost-sensitive multi-armed bandit data-sharing framework that requests data from multiple stakeholders. The hierarchical framework will extend the multi-armed bandit to multiple data sources and decompose similarity into two interconnected elements: manufacturer similarity and data similarity. In addition, the framework will assess and differentiate the contributions from multiple data owners in manufacturing based on the similarity metrics. In subsequent work, the PIs will extend the envisioned ecosystem to facilitate natural language queries, integration of manufacturer constraints, and novel data-sharing incentives. 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.