An intelligent neuro-system for failure detection and accommodation

Abstract

To enhance the performance of intelligent control systems, an automated, online procedure for observing changes in the dynamics of the controlled plant is needed. An interesting approach is the use of neural networks. A methodology using a neural network for failure detection and accommodation is presented. The main idea is to constantly monitor system output for off-nominal behavior (failures) and to use this information to generate an appropriate control action. A two-layer neural network is trained on input-output data pairs generated by simulating the system behavior in different failure modes. An integrated intelligent control system combining the controlled plant, a controller, a trained neural network for failure detection, a vector matching mechanism, and a neural network for failure accommodation is constructed. The vector matching mechanism cross correlates the output of the controlled plant with those of trained neural networks, and reports its decision about the system condition to a neuro-designer. The neuro-designer assesses the system dynamics and generates proper controller coefficients suitable for the current plant dynamics. The computed controller coefficients are continuously downloaded from the neuro-designer to the controller to ensure a stable operating mode and accommodate failures in the plant as they occur. A preliminary simulation result, conducted on the control of an airplane, showed that the intelligent controller is able to maintain system stability even in cases of harsh failures in a tilt-rotor airplane.