FPGA Based Design of Artificial Neural Processor Used for Wireless Sensor Network

Abstract

In this paper,  a simulation of  artificial intelligent system has been designed for processing  the incoming data of  sensor  units and then presenting proper decision. The Back-propagation Neural Network BPNN has been used as the proposed  intelligent system for this work, whereas the BPNN is considered as a trained network in conjunction with an optimization method for changing the weights and biases of the overall network. The main two features of the  BPNN are: high speed processing, and producing  lowest Mean-Square-Error MSE ( cost function ) in few iterations. The proposed BPNN has used the linear activation functions 'Satlins' and 'Satline' for the hidden and output layer respectively, and has used the training function 'Traingda' ( which is gradient descent with adaptive learning rate)  as a powerful learning method. It is worth to mention, that no previous research used these three functions together for such analysis. The MATLAB software package has been used for  designing and testing the proposed system. An optimal result has been obtained in this work, where the value of  Mean-Square-Error has reached to zero   in 87 epochs, and the real and desired outputs have been fitted. In fact, there is  no previous work has reached to this optimal result.  The proposed BPNN has been implemented in FPGA, which is fast, and low power tool.