Construction And Analysis Of Pulsed Neural Network Based On FPGA

With the increasingly serious adverse effects of electromagnetic interference on electronic systems,traditional protection methods have been difficult to protect electronic systems.It is of great significance to improve the reliability of electronic systems by drawing on the advantages of organisms in adaptive immunity and adopting new ideas to conduct research on electromagnetic interference resistance.At present,the implementation of pulse neural network based on software has slow processing speed,long time and large computer volume,and FPGA has the advantages of high parallelism,good real-time performance and small volume.Therefore,in the context of electromagnetic bionic protection,this paper implements the Izhikevich neuron model and the synaptic plasticity model of coexistence of excitatory synapses and inhibitory synapses based on FPGA,and constructs a feedforward pulse neural network with a hierarchical network topology.The anti-disturbance function of the network under impulsive noise,Gaussian noise and electric field noise is studied,which provides new ideas for electromagnetic bionic protection of electronic systems.The main work of this article is as follows:(1)Aiming at the realization of the Izhikevich neuron model,most of them use the Euler method to discretize,which has the problem of low accuracy.In this paper,the high-precision fourth-order Longge Kuta method is used,and the accuracy and algorithm of the Euler method are theoretically derived and algorithm verification.The state machine is used to implement the Izhikevich neuron model based on FPGA,and the software is used to fit the data and calculate the Euclidean distance.The results show that the fourth-order Longge Kuta method is more accurate than the Euler method.The FPGA-based Izhikevich neuron model is basically consistent with the software implementation data and the Euclidean distance is small and close to 0,which verifies the correctness of the FPGA implementation.Based on the FPGA implementation cycle of us level and software implementation of ms level,it shows that the FPGA implementation speed has increased by an order of magnitude.(2)For the realization of the synaptic plasticity model,most of them are excitable,and there is a problem of imperfect models.This article uses a more complete synaptic plasticity model in which excitatory and inhibitory synapses coexist.The division and exponential IP cores are called,and the STDP is implemented based on FPGA,and the data is fitted with the software implementation and the Euclidean distance is calculated.At the same time,a hardware description language was used to describe the synaptic plasticity model,and three Izhikevich neurons with synaptic plasticity were realized based on FPGA.The results show that the FPGA-based STDP and software implementation data are basically stable and the Euclidean distance is small and close to 0,which verifies the correctness of FPGA implementation.The regulation of synaptic plasticity of posterior neurons tends to synchronize with the firing of anterior neurons after a period of time.(3)For the anti-disturbance function of impulse neural network,this paper builds a random connection on FPGA to feed forward impulse neural network and imposes three types of noise interference,impulse noise,Gaussian noise and AC electric field.The influence of the correlation of neuronal membrane potential is compared with the anti-disturbance function of the fully connected feed-forward pulse neural network.The results show that under a certain intensity of three kinds of noise interference,the relative change rate of the discharge rate of the random feed-forward pulse neural network is small,and the correlation of the neuron membrane potential is high.The anti-jamming capability and anti-jamming function are better than the fully connected feed-forward pulse neural network.