Data-Driven Design of Distributed Kalman Filter-Based Fault Diagnosis for Large-Scale Systems

This article presents a data-driven distributed Kalman filter (DKF)-based fault diagnosis scheme, which can successfully detect actuator and sensor faults in large-scale dynamic systems, with heterogeneous subsystems interconnected through the directed topological graph. In the developed distributed framework, a local computing model (LCM) is formulated for each individual subsystem. In each LCM, based on the subspace identification method, the data-driven distributed Luenberger observer-based residual generator is constructed for each subsystem using local and neighboring process data. This offers an alternative expression of subsystem dynamics, supporting the design of more practical functions, for example, integrated monitoring and control design. The totally unknown interaction term is decoupled in the local residual, preventing the effect of actuator faults from propagating through the diagnostic network, so that the actuator fault isolation can be realized in each LCM. Then, the local noise covariance matrices are identified, and thus the data-driven DKF-based residual generator is formed, attaining improved detection performance by attenuating the effect of strong noises. Moreover, the adaptive configuration is developed for each subsystem, where the detector does not require retraining for changes in operating points or interconnection parameters, ensuring the continuity of the diagnostic process. Finally, case studies on the hot strip mill process verify the effectiveness and performance of the proposed method.