An Adaptive BP Neural Network Algorithm for 2-Joint Rigid Robots

According to the existing problems that back-propagation (BP) neural network with fixed activation function parameters has the shortcomings of slow learning speed, weak generalization ability and ease of falling into local minimum, an adaptive BP neural network algorithm for 2-joint rigid robot is proposed. Firstly, a novel structure of activation function with adjustable parameters is designed. In order to make the BP neural network have better nonlinear mapping capability, its mapping range, steep degree, and horizontal and vertical position parameters can be adjusted adaptively. Then, a fuzzy controller is designed to adjust the slope factor. Thus, the optimality of the slope factor can be guaranteed. Finally, a control system is designed and applied to the tracking control of a 2-joint rigid robotic system. The adaptive BP neural network algorithm is adopted to tune the proportional, integral and differential gains of the position controller. Simulation results show that in comparison with the classical BP neural network algorithm based on Sigmoid activation function with fixed parameters, the adaptive BP neural network algorithm takes the adaptability of activation function into account, improves the learning speed and generalization ability, and restrains false saturation. The convergence rate of position errors can be increased by 20 times, and the position and velocity tracking errors can be reduced to a small value close to zero with the proposed algorithm.