Pico-satellites represent the next logical step in the evolution of capable, low-cost satellite systems. Since these satellites are very small (less than 1 kg), they pose unique engineering challenges. Due to tight mass, size, and power restrictions, the attitude control of picosatellites is generally done by magnetic torquers. Reaction/momentum wheels are still under development to satisfy the mass and power requirements. As a consequence, attitude accuracy of picosatellites obtained is relatively low on the order of 10 degree. Typical missions involving imaging/remote sensing, in general, require high attitude accuracy on the order of 100th of a degree. Furthermore, in the cases of failures of attitude sensors and/or actuators, the satellite attitude drifts leading to loss of mission. In this paper, a reconfigurable intelligent attitude control system is proposed. The satellite may become dysfunctional due to the sensor/actuator failures. To overcome this problem and to ensure high fidelity, multiple sensors and actuators are implemented introducing high redundancy in the system which leads to increase in weight and power budget of the spacecraft. For the picosatellites where the tight mass and the low power restrictions are posed, implementing redundancy in sensors and actuators will increase the mass and the power requirements. In this paper, we suggest a novel approach based on reconfigurable controls. When an actuator/sensor fails, the attitude control capability reconfiguration is required. This demands that the control strategy must be adaptive to take care of the possible failures. This complicates the whole control design process besides making the system complicated. An intelligent control methodology based on neuro-fuzzy control is suggested in this paper. A fuzzy state noise driven extended Kalman filter based learning in dynamic neural network has been developed to provide highly adaptive and accurate neural control in real time. The dynamic neural network may not converge fast when using the backpropagation based learning scheme. Numerous literatures exist proving the superiority of extended Kalman filter based learning where in artificial noise is injected to ensure the positive definiteness of the state covariance matrix. However, because of finite precision, round-off errors and uncertainty in the a priori state covariance, filter may diverge or the convergence may be poor.
展开▼
