Hardware Implementation And Application Of BP Neural Network Based On FPGA

Artificial neural network can realize the non-linear mapping of arbitrary precision through the learning of the sample,so it is widely used in many fields such as system control,pattern recognition,signal processing,information prediction and so on.Under normal circumstances,neural networks are mostly implemented based on software.The neural network implemented by software cannot perform parallel operations and is limited in applications where real-time requirements are high.This paper presents a hardware implementation method of neural network based on FPGA.Due to the high-speed and parallel computing features of FPGA,the convergence speed of neural network is effectively improved.Firstly,the key problems of implementing BP neural network in FPGA are improved,and the bipolar S-type activation function of neural network is improved by using smooth interpolation method,and the resources of FPGA to realize the activation function are optimized.In the intermediate data storage,the use of a register circular buffer method is proposed,which solves the problem of insufficient external storage throughput,and ensures that parallel calculations can be performed within each module during the learning process.In the network construction,FPGA deep pipelining technology is used to implement serial and parallel networks,and the overall computing capability of the network is optimized.Secondly,using the hardware description language Verilog to implement the BP neural network modules and complete the overall architecture.The actual efficiency is compared with the PC platform through multiple test functions,and after 20 comparison experiments,It is proved that the speed of network convergence achieved by FPGA is beyond the software implementation speed by 3 orders of magnitude,which provides an important theoretical basis for the hardware implementation of artificial neural network.Finally,the prediction of the rolling force of the strip cold rolling process is the actual engineering background,and the design of the BP neural network rolling force forecasting model based on FPGA is completed.Use the data collected in the field to train and predict the network.The experimental results show that the accuracy of the model for rolling force forecasting is higher than that of the traditional rolling force forecasting model,and the training speed is much faster than using PC to realize the BP neural network rolling force forecasting model.