Research On Anti-radiation Interference Based On CNN Chip And SNN Chip Damage Analysis

Spatial radiation interference,especially the single even upset(SEU)effect,has a great impact on the normal and stable operation of the neural network chip.It will cause the bit parameters of the weight parameters stored in the chip SRAM memory to randomly flip,and then the weight parameters of the neuron The value will change,which will directly affect the accuracy of the neural network chip output.In recent years,with the development of artificial intelligence technology,the original micro control unit(MCU)has long been unable to meet the requirements of deep learning for massive data operations and high-speed operations,and artificial intelligence(AI)chips have emerged.For applications that are applied to chips in complex spaces,such as satellites,higher requirements are placed on the stable operation of the chips.This paper focuses on the research and analysis of the reinforcement technology for traditional chips,such as the method of hardware-reinforced three-mode redundant circuit,and the software reinforcement mainly includes the scrubbing method and error correction coding.Although these methods can improve the anti-jamming capability of the chip to a certain extent,for the neural network chip,a chip that emphasizes performance such as low hardware overhead,short recovery time,and fast processing speed,traditional reinforcement technology cannot play a very important role.Good function and traditional reinforcement technology also failed to make full use of the characteristics of the connection of each neuron of the neural network in the neural network chip.Based on the above problems of chip hardware overhead,recovery time and processing speed,it is proposed to use dropout algorithm to construct a new network framework,shield neurons affected by SEU with a certain probability,and carry out simulation verification of relevant experiments.The main work of this paper is as follows:1.This article makes a detailed study and analysis of the chip's space irradiation interference environment,analyzes the mechanism of the single event upset effect,and summarizes some existing chip anti-irradiation interference reinforcement technologies: process reinforcement,shield reinforcement and design The principle of reinforcement,and enumerate several common reinforcement methods such as three-mode redundancy,scrubbing methods and error correction coding principles and deficiencies in terms of commonly used design reinforcement methods.2.On the handwritten font data set,train convolutional neural network CNN and spiking neural network SNN to verify the accuracy of the undisturbed neural network's recognition of the data set from the accuracy.For the trained neural network,a network parameter and model structure are extracted,and the damage analysis of the SNN network is performed.3.The software simulates the effect of space irradiation interference,and randomly selects 1 ‰,1%,5%,and 10% of the proportional parameters for random bit error injection interference,so that a part of the wrong parameters can be obtained and the wrong parameters can be obtained.Replace the original parameters,and return to the corresponding layer of the network,assign the weight represented by the corresponding neuron,combined with the dropout algorithm and CNN network,test the accuracy of the neural network chip in the interference of irradiation,and verify the feasibility of the algorithm.