Optimization,Nonlinear Modelling And Control Of The Magneto-Rheological Actuator Of Robots

The service robots and the special robots are required to motion in the unstructured environments,e.g.,mountains,forests and homes;to provide safe and friendly services,such as grasping,hammering,etc.Therefore,the robots are demanded to be excellently compliant to ensure the safety of humans,the environments,the objects as well as themselves.Meanwhile,the robots should have a large load capacity and accurate force control accuracy for perfectly complete more jobs.The magneto-rheological actuator(MRA)introduces excellent compliance,large load capacity and good torque control accuracy.It is ideal for the service robots and the special robots that of large load capacity and frequently physical interacts with the outside world.It is therefore very significant in academic and engineer to study the accurate and rapid design methods and the high-precision force control method of the MRA.The magneto-rheological clutch(MRC)is the key module of the MRA.The very key point of its design is to determine the sizes of every components,the coils number and the magnitude of the input current to create the internal nonlinear magnetic field that generate a maximum clutch torque in a long-term usage.The finite element analysis(FEA)method is a “black-box" method that demands a great deal of computation resources.The empirical design skills of the software in engineer can greatly affect the design precision of the MRC,which hiders the rapid analysis and design of the MRC.A more efficient way is to use the magnetic equivalent circuit(MEC)method which,however,can only provide a ‘dirty’ model of the average value of the nonlinear magnetic field inside the MRC.The design by using the MEC method is also unreliable since the thermotics of the MRC can not be considered within this method.In this paper,based on the traditional MEC method,a nonlinear MEC method of the MRC is proposed by partitionally modelling the MRC according to the distribution characteristics and nonlinearity of the internal nonlinear magnetic field and solving the magneto-motive force of the nonlinear magnetic field.The proposed nonlinear MEC method provides the closest modelling precision to the FEA method but with much more computational efficiency.By founding out the conduction and radiation paths of the heat generated by the coils,the thermal model of the MRC is proposed to compute the temperature of its every components.Based on the two proposed model and with considering some practical engineering constraints,the mechanics-electrics-magnetics-thermotics optimal design method is proposed.Then the genetic algorithm is used to solved out the optics sides of every components,the coils number and the input current magnitude of the MRC.The severe rate-dependent hysteresis and creep nonlinearity between the input current and the clutch torque of the MRC leads to one input current corresponds to unlimited possible clutch torque,which introduces a great difficulty to the MRA torque control.In this paper,a magnetic sensor is embedded inside the MRC to study the influencers of the rate-dependent hysteresis and the creep.A sensor model and a torque estimation method based on the sensor value are proposed to estimate the real magnetic field and the clutch torque of the MRC,without requiring any data tests on the MRC prototype.For better torque estimation accuracy,the neural networks is adopted to model the ratedependent hysteresis of the MRC.Different from the “black-box” modelling approach as in the work of the predecessors,the diagonal recurrent neural network(d RNN)of the classical Preisach model is derived,and the feasibility,the convergence,the symmetry and the rate-dependency of the d RNN are deeply studied.The “white-box” approach of the d RNN in hysteresis modelling is therefore proposed with specifying the influence and functionality of each network parameters to the hysteresis modelling.A new loss function with considering the feasibility conditions is then proposed to force every neuron to model the rate-dependent hysteresis of the MRC in a stable manner.Address on the invalidity in modelling the creep nonlinearity and the large computation of the d RNN,a fractional multi-state differential model is proposed to efficiently and simultaneously model the rate-dependent hysteresis and creep nonlinearity of the MRC by introducing the torque saturation function and the fractional calculus.In order to accurately control the MRA output torque in a low-cost way,the model reference PID feedforward control method is proposed based on the fractional multi-state differential model.Address on the low control bandwidth and accuracy of the model reference PID feedforward control due to the rate-dependent hysteresis,the d RNN based inverse multiplicative structure feedforward control strategy is proposed.By optimally selecting the best neuron to minimize the feedforward control error,it provides an openloop MRC torque control of 97% precision and double high control bandwidth and control accuracy than the model reference PID control.To compensate the model uncertainties due to the inaccurate identification of the model parameters,a sliding-mode observer is proposed to estimate and compensate the modelling error of the multi-state differential model based on the super-twisting algorithm(STA).A STA based sliding-mode controller is further proposed to accurately control the MRC output torque in close-loop.Note that the MRA output torque is simultaneously determined by the motor speed and the MRC clutch torque,which is a typical two-input-one-output system.The dynamic model the MRA is therefore proposed and a speed feedback torque control law is established to control the MRA output torque.In the end,the MRA is applied to develop a compliance and heavy-load leg mechanism for the quadruped robot,which is capable to carry weights up to 50 kg and provides accurate force control and great compliance control.