Design Of Soft Error Excitation System For Fault Tolerance Research Of Visual Convolutional Neural Network

The research on fault tolerance of convolutional neural networks is to build a new and more reliable network model.True fault tolerance refers to the ability of the neural network to maintain normal application functions when an abnormality occurs in the system architecture layer.In the field of aerospace,soft errors caused by the space radiation environment will bring hidden dangers to the stable operation of the system and the application of convolutional neural networks.Therefore,research on fault tolerance of convolutional neural networks based on soft errors is very important.This thesis will focus on this aspect to design a soft error excitation system suitable for the fault tolerance research of visual convolutional neural networks.This thesis models the key components of the system architecture layer and the upper-layer convolutional neural network application on the system-level simulation platform.By simulates the generation of soft errors through fault injection,this paper constructes a soft error generation system based on fault injection.At the same time,this thesis designs a data collection scheme for the system architecture layer and the application layer in order to form an automated soft error occurrence and data sample collection architecture.To obtain the path of fault propagation from the architecture layer to the program layer,this thesis uses a two-state Bayesian discrete network to perform hierarchical cascade analysis on the observation nodes of the system architecture layer and the application layer,and after learning through different hierarchical structures and parameters,the Bayesian network topology and Bayesian conditional probability table are obtained.The forward reasoning of the topological graph can detect the accuracy of the Bayesian network,and the reverse reasoning of it can be concluded that in the case of silent data corruption(SDC)failure,the key register or convolutional neural network level that is most likely to be abnormal.In order to make the fault performance set of the convolutional neural network more complete,this thesis designs a soft error fault excitation scheme to obtain more fault samples through Bayesian decentralized fault injection.Based on the Bayesian weight,the ratio of the number of injections occupied by each register is adjusted to make the number of fault experiments.Under the same condition and consistent time overhead,the 3.33% SDC failure rate of the sparse injection was increased to 6.11%.