A New Sensor Fault Diagnosis Technique Based Upon Subspace Identification and Residual Filtering

This paper presents a new methodology for designing a detection, isolation, and identification scheme for sensor faults in linear timevarying systems. Practically important is that the proposed methodology is constructed on the basis of historical data and does not require a priori information to isolate and identify sensor faults. This is achieved by identifying a state space model and designing a fault isolation and identification filter. To address time-varying process behavior, the state space model and fault reconstruction filter are updated using a two-time-scale approach. Fault identification takes place at a higher frequency than the adaptation of the monitoring scheme. To demonstrate the utility of the new scheme, the paper evaluates its performance using simulations of a LTI system and a chemical process with time-varying parameters and industrial data from a debutanizer and a melter process.