单粒子翻转对神经网络的影响分析与优化

DNN芯片作为星载芯片应用到卫星系统中时会受到太空辐照的影响，其中单粒子翻转对存储单元的干扰会使得存储器单元的参数出现错误，该错误映射到神经网络中会造成神经网络最后的输出结果出现偏差。结合单粒子翻转概率模型，对用于网络推断的神经网络的权值参数进行注错后分析实验结果准确率，从激活函数的非线性特性分析并通过实验验证具有双边抑制效果的函数容错能力更强。进一步在网络卷积层后加入BN层和在训练过程中考虑L2正则化提高网络的容错能力，并通过实验验证其可行性。

Analysis and optimization of single event upset on neural network

When the DNN (deep neural network) chip is used in a satellite system as a space-borne chip, it will be affected by space radiation. The interference of single event upset (SEU) on the storage unit will cause the parameters of the memory unit to be wrong. The error mapping to the neural network will affect the output results. This paper analyzes the accuracy of the network inference which combines the SEU probability model to inject the error on the network weight parameters. From the analysis of the nonlinear characteristics of the activation function and experimental verification, it is found that the activation function with bilateral inhibition is more fault-tolerant. Furthermore, we add the BN layer after the network convolution layer and consider L2 regularization during training to improve the network's fault tolerance, and verify its feasibility through experiments.