Direct identification of fault estimation filter for sensor faults

Abstract: We propose a systematic method to directly identify a sensor fault estimation filter from plant input/output data collected under fault-free condition. This problem is challenging, especially when omitting the step of building an explicit state-space plant model in data-driven design, because the inverse of the underlying plant dynamics is required and needs to be stable. We show that it is possible to address this problem by relying on a system-inversion-based fault estimation filter that is parameterized using identified Markov parameters. Our novel data-driven approach improves estimation performance by avoiding the propagation of model reduction errors originating from identification of the state-space plant model into the designed filter. Furthermore, it allows additional design freedom to stabilize the obtained filter under the same stabilizability condition as the existing model-based system inversion. This crucial property enables its application to sensor faults in unstable plants, where existing data-driven filter designs could not be applied so far due to the lack of such stability guarantees (even after stabilizing the closed-loop system). A numerical simulation example of sensor faults in an unstable aircraft system illustrates the effectiveness of the proposed new method.