This work addresses the problem of detecting parametric faults in nonlinear dynamic systems by extending an eigenstructure based technique to a nonlinear context. Two local state-space models are updated online based on a recursive subspace system identification technique. One of the models relies on input-output real-time data collected from the plant, while the other is updated using data generated by a neural network predictor, describing the nonlinear plant behaviour in fault-free conditions. Parametric faults symptoms are generated based on eigenvalues residuals associated with two linear state-space model approximators. The feasibility and effectiveness of the proposed framework are demonstrated through two case studies.
- Feedforward neural network
- Model-based fault detection
- Parametric fault
- Recursive estimation
- Subspace system identification