Neuro-fuzzy architecture of the 3D model of massive parallel actuators

Pasila, Felix and Alimin, Roche and Natalius, Hans (2015) Neuro-fuzzy architecture of the 3D model of massive parallel actuators. ARPN, Volume (12). pp. 2900-2905. ISSN 1819-6608

PDF Download (187Kb)

Abstract

This paper reports the effective control mechanism of the discrete state manipulators (DSMs) with six degree of freedom (DOF). The DSMs are special kind of robot manipulator with massive actuators that can be switched among limited number of discrete states. We introduce ternary-DSMs (t-DSMs) manipulators which controlled by force and have continuous motions that commanded through only three discrete states. The main problem of this mechanism is how to design a real-time controller which is efficient and fast for solving its inverse static problem (ISP). Precisely, a computational intelligence method based on neuro-fuzzy method is suggested to find the optimal training computation, which is measured by the root mean squared error of ISP. The architecture of t-DSMs featuring three-state force pneumatic actuators and six-DOF. For instance, a neuro-fuzzy method for t-DSMs constructs IF-THEN rules from fuzzy relations among inputs and outputs in the training mechanism (inputs: position and force outputs: three-state). After related model is found in the training phase, the architecture can be used to determine outputs of the network from given inputs with similar accuracy in the testing phase. The paper proposed an architecture which is based on the Neuro-Fuzzy Takagi Sugeno (NFTS) inference scheme with Gaussian membership functions. The structure is with multivariate input and multistate outputs, such as positions and forces as input NFTS networks and the three-state of the actuators as output networks. The learning of the network uses an extended LMA version with optimal training parameters. The training algorithm needs at least one million iterations with different membership functions employ around 17 of the input-output correspondences dataset from the known input and output. For training database, the NFTS model generates 124 dataset from the 729 possible dataset. The optimized membership function (M) after one week searching time using optimized search procedure using M from 4 to 15 for the 6-DOF model of 6-ternary DSMs. Regarding model performances for the ISP solution, the NFTS with M=9 features better root mean squared error results compared to the others.

Actions (login required)