Neural fault detection, isolation, and estimation design

For autonomous and high performance systems, onboard fault detection, isolation, and estimation (FDIE) against sensor, actuator, and system failures are important features in ensuring performance and avoiding catastrophes. Existing FDIE designs can be roughly classified as trial-and-error, hardware redundancy, model-based analytic redundancy, or knowledge based expert system approaches. All are subject to, fully or partly, uncertainty, nonlinearity, and complexity, leading to high false alarm rate or erroneous classification. In this paper, a neural FDIE scheme is developed and verified using simulated data. The neural FDIE scheme has the following advantages: robustness against unmodeled dynamics; ability in handling nonlinear dynamics; flexibility in accounting for both discrete (jump) behavior and continuous degradation; modular and systematic design; and potential for unanticipated failures. The FDIE scheme is based on the general regression neural network (GRNN) concept, making the FDIE easy and fast to train. The design is verified using an F/A-18 system. A class of failure patterns are simulated. The performance and properties of the neural FDIE scheme are assessed.