Neuro-Super Twisting Sliding Mode Control of a Reaction Wheel

Abstract

Sliding mode controllers are renowned for their robustness and fast dynamic response in speed control of direct current motors. The major drawback to their application is the chattering phenomenon which is harmful to actuators like the reaction wheel and causes performance degradation. To reduce this effect, the super twisting sliding mode control is employed because it tends to resist varying load parameters and reduce chattering. However, the super twisting SMC has two gain pairs that need to be tuned via the trial-and-error method to attain the optimum performance of the controller. This process of trial-and-error tuning makes the implementation of the super twisting SMC hectic and stressful. Therefore, an adaptive method of tuning these gain pairs is proposed in this work that can automatically tune the gain pairs to the required optimum value that matches the change in varying parameters, thereby making it robust to changes. This method is achieved by designing, training, and implementing a shallow neural network called fitnet, which is a multilayer perceptron, to effectively fit the sliding variable to the target curve in finite time. In doing so, the neural network targets the nominal control taking input from the sliding variable and its delayed data, and then predicts the appropriate nominal control input needed to keep the system states state trajectories on the sliding surface in finite subsequent time. More so, this work was simulated using MATLAB/Simulink and a corresponding enhanced performance objective (rise time of 0.7471s, settling time of 1.3493, peak value of 1.001, zero percentage overshoot and undershoot, and steady-state error of 5.951e-05) over the PID, conventional SMC, and super twisting SMC were recorded respectively. In addition, the proposed neuro-tuned SMC gave an improved control performance when compared with the work of Rakhonde and Kulkarni and Morfin et al