A Distributed Data-Driven Projection-Based Fault Detection Scheme for Large-Scale Dynamic Systems

This article presents a distributed data-driven fault detection scheme for large-scale uncertain dynamic systems with sensor networks. To this end, a data-driven orthogonal projection-based fault detection method is first developed in the centralized handling fashion. Specifically, the system nominal dynamics are modeled in the kernel subspace within a Hilbert space. Process data are projected onto the orthogonal complement of the kernel subspace, serving as the basis for generating residuals in uncertain systems. An adaptive threshold is incorporated to reduce diagnostic conservatism. Given the distributed architecture of sensor networks in large-scale systems, a consensus-based approach is employed to realize the proposed method in the distributed fashion. The average consensus algorithm enables projection-based residual generation and evaluation functions to be performed at each node, while achieving identical detection performance with the centralized scheme. The effectiveness of the proposed scheme is demonstrated through a case study on finishing mill in a hot rolling process.