Neural Network Based Fault Detection for Fault Tolerant Control of Systems with Multiple Magnetic Actuators and Sensors

There is considerable interest in the improvement of fault tolerance in the design of active position and vibration control systems. In the creation of many magnetic actuator position or vibration control systems there exists potential for the inclusion of redundancy in the number of actuators and sensors used for control. This redundancy can provide improved tolerance to actuator or sensor related faults if, when a fault occurs, control of the system can be rapidly reconfigured to bypass control from a faulty component to the remaining healthy ones. To do this requires a system for the detection and identification of faults as and when they occur. The performance of modern computerised control hardware is now sufficient to allow such a system to run in real-time, in parallel to any digital control algorithm. In this paper the development of a method for the detection and isolation of faults relating to control sensors and actuators is presented as a basis for the implementation of a fault tolerant control scheme. The method is based on the use of a neural network, operating in real time, for the detection of signal errors occurring in the plant inputs or outputs. The neural network is trained off-line using identification data taken from the plant and therefore does not require an accurate model of the plant dynamics. Results are presented for the application of this method to an active magnetic beating/rotor system, both in simulation and experiment. The issues of sensitivity to faults, speed of response and the effect of external disturbances on the reliability of the fault detection system are investigated and discussed. It is demonstrated that, through the implementation of a reconfigurable control scheme with a fault detection system, improved tolerance to sensor and actuator related faults can be achieved.