A Neural Narx Network Controller for a Pneumatic Artificial Muscle Driven Translational System
Electric motors and hydraulic/pneumatic cylinders are the most common manipulators used in robotics, prosthetics, and other mechatronic applications. All these types of actuators have their own advantages and limitations. Recently, the pneumatic artificial muscle (PAM), sometimes known as a McKibben actuator, has become a popular choice for manipulation in robotics and industrial applications. These actuators consist of a stiff, reinforcing braided mesh of fibers surrounding an elastomeric bladder (usually made of rubber), and two airtight end fittings. The rubber bladder swells as a pressurized internal fluid is introduced into it, and the constrained length of the double helix stiff fibers cause contraction in the axial direction. Either air or liquid can be used as the internal fluid. Compressibility and high flowability of air, together with the compliant behavior of the rubber bladder, can render a smooth, jerk-free, antagonistic operation of the PAM. The inextensible fiber is usually made of metal, nylon, carbon fiber, glass fiber, or other synthetic high-strength fibers. With commercially available tubing and sleeve material, the manufacturing cost of PAMs is reduced. The advantages of high power/weight and power/volume ratios, added safety, and operator ease-of-use (due to compliance and variable stiffness properties) make the PAM useful for a wide range of applications, including prosthetics, orthotics, humanoid robots, and morphing wing technologies. Although PAMs have many advantages over other type of actuators, it is still not a fully developed technology. The nonlinear characteristics of the dynamic behavior of PAMs limit the controllability, whereas air compressibility and the lack of damping ability causes dynamic delay of the pressure response and subsequent oscillatory motion.
An antagonistic tendon driven PAM system is developed in this research. The system has a single degree-of-freedom translational movement with pressure variation in the antagonistic PAMs. A slider moves over a guide rail with the pulling action of two PAMs in antagonistic fashion. A binary (on/off) control pneumatic solenoid valve is used for compressed air handling in the PAM. The overall experimental setup is inexpensive, due to the use of a low-cost solenoid valve and the custom assembled PAMs. A webcam tracks the position of a red dot on the slider for the slider displacement measurement with a low error tolerance limit (approx. 0.3 mm). A nonlinear autoregressive network with exogenous input (NARX) model of the relationship between the slider displacement and the valve control signal is developed using experimental training data. A neural NARX network can predict a dynamic time series output signal based on previous values of the output signal and previous values of an external signal. Two delayed displacement measurements are used as external inputs, and two delayed control signals are used as feedback into the neural NARX network. The network has a single hidden layer with five neurons. Some randomly generated data is used as the training data for the neural NARX network. The network parameters are updated through training with the Levenberg-Marquardt backpropagation method. The trained network can predict the required control signal in the next time step to generate a predefined target slider displacement trajectory. The neural NARX model is effective in recognizing the highly nonlinear behavior of this antagonistic PAM driven system. The model can predict the required control signal from the reference slider displacement pattern. The regenerated slider displacement trajectory closely follows the reference profile. Thus, a shallow neural NARX network model of the PAM dynamics can avoid the problem of modeling the nonlinearity, while still controlling the system.
A Neural Narx Network Controller for a Pneumatic Artificial Muscle Driven Translational System
Category
Poster Presentation
Description
Session: 17-01-01 Research Posters - On Demand
ASME Paper Number: IMECE2020-25157
Session Start Time: ,
Presenting Author: Tushar Mollik
Presenting Author Bio: Tushar Mollik is a graduate student at Mechanical and Aerospace Engineering department at North Carolina State University. He is currently pursuing a PhD degree in Mechanical Engineering. His research interest is in soft robotics, machine learning, and control.
Authors: Tushar Mollik North Carolina State University
Edmon Perkins North Carolina State University