According to the White House’s 2022 National Strategy for Advanced Manufacturing, building resilience into manufacturing supply chains is a top priority – one that has only grown in importance. This directly refers to manufacturing networks in which interconnected manufacturers (nodes) exchange data and other sensitive and competitive information.
To improve resiliency and other metrics, incorporating artificial intelligence (AI) and machine learning (ML) is a promising technology for today’s manufacturing networks. However, there is an increasing divide between large manufacturers and small and medium manufacturers (SMMs) in adopting digital tools and leveraging their benefits on resilience.
Mechanical Engineering Professor Thor Wuest is currently leading a project that aims to close that gap and ensure an even-level manufacturing network for SMMs – the backbone of the manufacturing industry.
Wuest’s two-year, nearly $300,000 National Science Foundation-funded project aims to develop a private blockchain-enabled federated learning framework. His post doc Mojjtaba Farahani and Ph.D. student Md Irfan Khan are working on the project.
“Smaller companies often struggle with demands from their customers, which are often large manufacturers, to prove more data needed for forecasting delivery changes, lead time or any interruption within the supply chain,” Wuest says. “The original equipment manufacturer (OEM) wants to know if there are any delays in the supply chain or if a problem can be anticipated to improve their operations and overall resiliency.”
The COVID-19 pandemic showed the vulnerabilities of manufacturing networks. AI and ML may eventually solve supply chain issues, but they have not been implemented at scale throughout manufacturing networks extensively. This is especially true with SMMs due to their limited resources and smaller IT departments, which makes them unable to participate in many high-value manufacturing networks that often require specific technologies and data sharing.
“We wanted to develop a framework where the burden of implementing these technologies would be on the larger companies who profit the most from their insights while ensuring SMMs have full control over their data and what to share,” Wuest says.
The problem stems from manufacturing networks inability to securely and efficiently exchange data and leverage network level learning due to cybersecurity concerns and a hesitation to share private data. Using federated learning, AI models can be collaboratively trained on decentralized nodes by only sharing model parameters without disclosing the underlying data. Wuest’s project aims to enhance the resiliency of SMMs by providing access to a secure, private blockchain network that enables transparent communication while preserving the confidentiality of proprietary competitive (raw) data and locally trained ML models.
"Manufacturing industries are in dire need of advanced AI and ML platforms to accelerate their transition toward Industry 4.0 and smart manufacturing systems,” Farahani says. “As researchers, it is our responsibility to translate innovation into practical tools. At our lab, we are committed to pioneering these efforts, to ensure manufacturers of all sizes remain competitive and at the cutting edge of technology.”
Federated learning is a distributed ML approach that enables collaborative model training across multiple servers. Instead of sharing raw data, each participant trains a local model and only shares the updates. These updates are then aggregated to build a global model, which helps optimize and enhance the performance of local models. Federated learning plays a key role in advancing intelligent and efficient manufacturing ecosystems by enabling secure, confidential, and decentralized collaboration.
This will open new business opportunities as individuals or companies can invest in small machines and start generating valuable data.
- Md Irfan Khan
The project also aims to integrate a private blockchain into the federated learning framework to manage access controls and model updates. Unlike existing approaches, the framework will focus on specific challenges and requirements of manufacturing networks by utilizing algorithms for coordinating models. This includes ensuring confidential data remains local under full control of the individual nodes and reducing overhead costs for SMMs.
“The blockchain architecture would be developed by the OEM, but the SMM can directly communicate with other nodes through network channels using their private keys,” Wuest says. “We wanted to see each node train their own model with their own data and then release the model parameters to build that global model that then goes back and helps them with their own initiative.”
Blockchain technology enables secure tracking of assets and transactions across distributed networks and is particularly promising for improving transparency and resilience in manufacturing networks. Based on Wuest’s framework, even if model parameters are leaked, they are less sensitive than raw production data since they cannot be directly attributed to specific manufacturing processes. If model parameters are tampered, blockchain’s immutable ledger provides traceability to help identify the source and nature of the breach.
“At this stage, I’m running simulations using benchmark time series datasets. If successful, the next step will be setting up multiple edge servers, where each server represents a different industry,” Khan says. “The setup will help us evaluate the performance of the federated learning method more realistically, and we’ll see how they perform before implementing the blockchain. The healthcare industry is already leveraging this approach due to the sensitivity of patient data.”
Small manufacturers often operated with limited data due to smaller batch sizes, making it difficult to build robust ML models. Federated learning can help address data scarcity, protect privacy and maintain competitiveness by enabling collaborative model training across multiple small organizations.
“There are similarities when a problem is based on data scarcity, privacy and the need to share,” Wuest says. “For example, a heart condition and a manufacturing tool may have different parameters but face a similar problem of small data size and privacy.”
According to Khan, federated learning can be applied to this project since all the suppliers can share the model parameters after training locally and the sensitive data remains secure.
“The federated learning mechanism is aggregating the model parameters and making a global model so they [SMMs] can benefit from all the data sets that all companies have,” Khan says. “By integrating blockchain, we can ensure transparency in model updates and safeguard the system against tampering or unauthorized access.”
“For example, a company manufactures 1,000 engine blocks as a basis for the data set. They don't want to share that with anybody because it’s competitive and financially sensitive information. But they could train a local model to predict tool wear, share the model parameters and the insights can help other companies,” Wuest says.
Data sets in manufacturing supply chains have multiple components and sharing them can be a challenge. But model parameters are a much lighter weight to share.
“The models can be built locally, which gives companies an incentive because they’ll use it to improve their own production. There's already an incentive for quality control on the local level that, by association, will feed into the global model,” Wuest says.
While limited research exists on how federated learning can enable decentralized machine learning in industrial systems and manufacturing networks, Wuest believes opportunities exist to integrate federated learning into automated entities for continuous model improvement.
“This will open new business opportunities as individuals or companies can invest in small machines and start generating valuable data,” Khan says. “In industrial settings, failure data is often rare and difficult to obtain. We can intentionally generate failure scenarios in a controlled lab setup to enrich the dataset. These insights can then be shared through a global federated model, allowing other participants to benefit from improved performance.”