Intelligent model-based control of complex multi-link mechanisms

Complex under-actuated multilink mechanism involves a system whose number of control inputs is smaller than the dimension of the configuration space. The ability to control such a system through the manipulation of its natural dynamics would allow for the design of more energy-efficient machines with the ability to achieve smooth motions similar to those found in the natural world. This research aims to understand the complex nature of the Robogymnast, a triple link underactuated pendulum built at Cardiff University with the purpose of studying the behaviour of non-linear systems and understanding the challenges in developing its control system. A mathematical model of the robot was derived from the Euler-Lagrange equations. The design of the control system was based on the discrete-time linear model around the downward position and a sampling time of 2.5 milliseconds. Firstly, Invasive Weed Optimization (IWO) was used to optimize the swing-up motion of the robot by determining the optimum values of parameters that control the input signals of the Robogymnast’s two motors. The values obtained from IWO were then applied to both simulation and experiment. The results showed that the swing-up motion of the Robogymnast from the stable downward position to the inverted configuration to be successfully achieved. Secondly, due to the complex nature and nonlinearity of the Robogymnast, a novel approach of modelling the Robogymnast using a multi-layered Elman neural ii network (ENN) was proposed. The ENN model was then tested with various inputs and its output were analysed. The results showed that the ENN model to be capable of providing a better representation of the actual system compared to the mathematical model. Thirdly, IWO is used to investigate the optimum Q values of the Linear Quadratic Regulator (LQR) for inverted balance control of the Robogymnast. IWO was used to obtain the optimal Q values required by the LQR to maintain the Robogymnast in an upright configuration. Two fitness criteria were investigated: cost function J and settling time T. A controller was developed using values obtained from each fitness criteria. The results showed that LQRT performed faster but LQRJ was capable of stabilizing the Robogymnast from larger deflection angles. Finally, fitness criteria J and T were used simultaneously to obtain the optimal Q values for the LQR. For this purpose, two multi-objective optimization methods based on the IWO, namely the Weighted Criteria Method IWO (WCMIWO) and the Fuzzy Logic IWO Hybrid (FLIWOH) were developed. Two LQR controllers were first developed using the parameters obtained from the two optimization methods. The same process was then repeated with disturbance applied to the Robogymnast states to develop another two LQR controllers. The response of the controllers was then tested in different scenarios using simulation and their performance was evaluated. The results showed that all four controllers were able to balance the Robogymnast with the fastest settling time achieved by WMCIWO with disturbance followed by in the ascending order: FLIWOH with disturbance, FLIWOH, and WCMIWO