High-Dimensional Cross-Plant Process Monitoring With Data Privacy: A Federated Hierarchical Sparse PCA Approach

Modern large-scale manufacturing systems typically involve joint efforts of multiple plants to finish the final product. The advances of information technologies, on the other hand, enable the collection of massive process variables in each plant that are critical to product quality. Monitoring such high-dimensional cross-plant processes is promising for achieving a system-level quality assurance. However, the increasing importance of data privacy prevents the fusion of data from different plants for a centralized process monitoring. To address this dilemma, this article proposes a novel privacy-preserving cross-plant process monitoring framework by leveraging a federated learning (FL) concept. Specifically, a new principal component analysis (PCA) with a distributed structure and a hierarchical sparsity regularization is first developed to reduce the data dimension. It can identify significant plants and variables in each eigenvector and, meanwhile, facilitate an efficient FL-embedded optimization algorithm to exchange model parameters, rather than raw data, between different plants. The local data are, thus, protected as they never leave their dwelling plant and cannot be inferred by others either. A control chart is then built based on the PCA outputs to jointly monitor all the plants using a safe FL protocol. The superiority of our proposed process monitoring framework is validated extensively by numerical simulations and two real industrial case studies.

Automated summary

This article proposes a novel privacy-preserving cross-plant process monitoring framework using federated learning. It reduces the data dimension with a new distributed principal component analysis (PCA) method and exchanges model parameters between different plants using a secure federated learning protocol. The superiority of this framework is validated through numerical simulations and real industrial case studies.