Design Considerations and Guidelines
The digitization of the manufacturing sector has enabled the availability of large volumes of data , this allows industry to embrace machine learning ( ML ) algorithms to identify and analyze patterns in the data . These patterns can be leveraged to enable decision support for activities including process and production planning , asset performance management and energy efficient manufacturing practices etc . The accuracy and robustness of ML algorithms depend heavily upon the amount , variety , veracity of training data .
To this end , a data-sharing ecosystem is encouraged where organizations participate and contribute their data as a strategic resource for the benefit of all . The incentive for participating in such an ecosystem is the value generated by access to a larger pool of data for model training . However , due to industrial competition and data privacy concerns , organizations are reluctant to share potentially commercially sensitive data , thus , the datasets remain in silos .
Federated learning ( FedL ) offers a potential solution to address the conflict between data protection and participation in a data sharing ecosystem . FedL enables organizations to collaboratively train robust AI models , without the need to directly share sensitive data with each other . Despite several contributions in domains such as natural language processing and healthcare , multiple barriers exist that are hindering the uptake of FedL in the manufacturing industry .
A key challenge is the complexity associated with designing and deploying a FedL solution . It requires consideration of many constraints such as the application type , FedL client configuration , global model training orchestration , choice of encryption mechanisms to secure models and incentive mechanisms . Currently , there is a lack of a clear methodology that allows practitioners and industry stakeholders to identify and evaluate the potential of using a FedL approach for their specific use case scenarios .
While the literature explores each of the design constraints independently , there is a need to consolidate these into a common framework to support the development of FedL solutions for smart manufacturing .
The purpose of this paper is to translate practical experience of designing , building and deploying FedL solutions as well as an analysis of current literature into guidance that :
• provides a perspective on the business models that enable and foster a collaborative ecosystem to increase participation in data sharing arrangements that are critical to the development of FedL solutions .
IIC Journal of Innovation 19