A Nonlinear Hybrid Fault Detection, Isolation and Estimation using Bank of Neural Parameter Estimators

This paper presents two novel hybrid schemes for component fault diagnosis in a general nonlinear system. Unlike most fault diagnosis techniques, the proposed solution detects, isolates, and identifies the severity of faults in the system within a single integrated framework. The proposed technique is based on a bank of adaptive neural parameter estimators (NPE) and a set of parameterized fault models. At each instant of time, the NPEs provide estimates of the unknown fault parameters (FP) that are used for fault diagnosis. Two structures of NPE, namely series-parallel and parallel, are proposed with their respective fault isolation policies. While the series-parallel structure possesses fast convergence, the parallel scheme is extremely robust to measurement noise. Although, it has a more complex isolation policy, the parallel structure exhibits a more robust fault isolation capability. The parameter estimation for both architectures is based on an on-line minimization of instantaneous output estimation error. Simple network architecture and straightforward weight adaptation laws make our proposed technique appropriate for real-time implementation. Simulation results presented in this paper for detection, isolation, and identification of faults in nonlinear reaction wheel actuators of a 3-axis stabilized satellite in the presence of disturbances and noise demonstrate the effectiveness of our proposed fault diagnosis schemes.