Comparative Modeling Study of Pneumatic Artificial Muscle Using Neural Networks, ANFIS and Curve Fitting

Pneumatic Artificial Muscles (PAMs) are widely used in the fields of biorobots and medicine due to their flexibility, safe usage, lack of mechanical wear, low cost of manufacturing, and high ratio of power to weight. Obtaining an accurate PAM model is crucial for building a controller that obtains the required performance specifications. This study aims to create various models for a PAM and to evaluate them with respect to their accuracy in reflecting PAM behavior. An experimental-based modeling approach was adopted to collect the necessary data in order to accurately model the PAM. The data were collected for different pressure setpoints and with different loads. Four system modeling techniques were utilized: (i) curve/surface fitting, (ii) Multi-Layer Perceptron Neural Network (MLP NN), (iii) Nonlinear Auto-Regressive with eXogenous (NARX NN) and (IV) Adaptive Neuro Fuzzy Inference System (ANFIS). The analysis of the four developed models showed that the performance of the MLP NN model exceeded all other models by having the smallest error. Therefore, a simple feedforward neural network can represent the complex muscle system compared to other complex modeling techniques.