Research On Application Of ELM In Modeling And Control Of Robots

Modeling and control system is one of the most important research fields of robotics,which the industry, academic and others paid more attention to. The development of robotsmainly depends on the high level of the control technology. Due to many uncertainties such asnonlinearity, strong-coupling, time-variability and many uncertain factors, so an exact mecha-nism model is tough to be built in robot system. As the nonlinear system identification andcontrolling way relying less on the exact mathematic model, neural networks (NN) is widelyused. Extreme learning machine (ELM) is single-hidden layer feedforward neural (SLFN) isresearched in recent years, which has unique Network structure training algorithm can havefast training and learning speed(that is faster than thousand times of traditional networks)Which has strong ability of approximation (if hidden nodes are more enough it can close tozero error). which can overcome the complexity of traditional network training algorithm andbe easy to stuck in minimum localization. ELM can be easily used to address the dilemmas inthe past when applied to the nonlinear system identification and process. On basis ofcomparison between with the traditional method and ELM, the experiment presents highefficiency and reliable practicability of this method. The research of thesis is sprawled out asfollows:(1) Through analyzing the basic characteristics and learning algorithm of ELM, furtherstudy its learning algorithms, we can get the offline and based on recursive least squares (RLS)online sequential extreme learning machine (OS-ELM)’s learning algorithm. Deducing indetail complex robot lagrangian dynamics mathematic model as basic theory of the nonlinearsystem identification and control.(2) The thesis studied nonliear auto regressive moving average (NARMA) system,multidimensional nonlinear system, hydraulic drive robot arm, golf robot and seven degree offreedom SARCOS anthropomorphic robot arm in the application of system identification,aiming at different system has single input single output or multiple input multiple outputdesigned the different identificational network structure of ELM. Experimental results showthat the ELM not only has fast and stable learning speed but also has high identificationaccuracy compared with radial basis function (RBF) and support vector machines (SVM) inthe same conditions.(3) The thesis studied ELM in the application of controlling. Firstly, aiming a traditionalnonlinear system made a trajectory control. Secondly, basing on ELM designed a compensa-tion controller of single-arm robot to realize the error compensation of manipulator position.Lastly, basing on the ELM in the application of a two degrees of freedom of the manipulatortrajectory tracking controlling, using inverse model of uncertain part of manipulator designed the ELM controller. The simulation results show, comparing with other networks, ELM canget a small control error and a good trajectory tracking performance.