Neural network chip anti-radiation system and method for space application

• Main Authors: XIE ZHUOCHEN, LIANG XUWEN, QIAN HUAN, WANG HUILING
• Format: Patent
• Language: Chinese; English
• Published: 26.06.2020
• Subjects: CALCULATING; COMPUTER SYSTEMS BASED ON SPECIFIC COMPUTATIONAL MODELS; COMPUTING; COUNTING; PHYSICS

The invention discloses a neural network chip anti-radiation system and method for space application. The system comprises a neural network chip and an FPGA chip. For a neural network chip, a method for withdrawing net is adopted, some neurons are discarded at a certain probability in the network training process; then, in the process of configuring the neural network into a neural network chip tocarry out network inference, error parameters caused by single event upset are mistakenly positioned, set to be zero and rewritten into a weight storage area of the neural network chip, so that weight parameter refreshing work of the neural network chip is realized, the influence of the error parameters on the accuracy of the whole network is reduced, and an anti-radiation function is further realized. 本发明公开一种针对空间应用的神经网络芯片抗辐照系统及方法，包括神经网络芯片及FPGA芯片。针对神经网络芯片，采用退网的方法，在网络训练过程中以一定概率丢弃一些神经元，然后，在将神经网络配置到神经网络芯片中进行网络推断的过程中，通过对单粒子翻转造成的错误参数进行错误定位并将其置0，并重新写入神经网络芯片的权值存储区，实现神经网络芯片的权值参数刷新工作，减少错误参数对整个网络的准确率的影响，进而实现抗辐照功能。