Research On Fault Prediction And Diagnosis Method Of Processor System Based On Fault Injection Simulation

In the field of aerospace,on-board systems with high operational stability are used for program control,data transmission and state monitoring.However,radiation factors and high-energy electrons in the space environment will bring hidden dangers to the stable operation of the system.Therefore,early health assessments of on-board systems are particularly important.The health assessment of on-board systems consists of two aspects: simulated fault injection to form sample datasets and complex system fault prediction,diagnosis.This article will focus on these two aspects,and research on the complex processor systems based of the ARM architecture in the on-board system.A hierarchical system-level model of key hardware components and upper applications of typical processor systems is built on Simics,which is a system-level simulation development platform,and a target system for fast evaluation of fault injection effect is obtained.On this basis,the processor micro-architecture and disk registers in the hardware layer perform singlebit flip fault injection,and the fixed timing sampling of system key variables is performed according to the execution time of the application.In this process,the timing control task is used to automate fault injection and timing acquisition,which reduces the time overhead caused by manual operation.According to the characteristics of time series data of system key variables acquired by the acquisition,a specific data cleaning method such as data standardization and failure mode labeling is designed,and a structured fault sample data set composed of {fault source,observation point uncertainty,and failure label} is constructed.Using the Bayesian network model in the graph model,the fault propagation relationship network in the data-driven processor system is established.The network topology is defined in a semi-automatic way,and the conditional probability table of the network is trained by means of sample learning to realize the bidirectional probability inference between "fault source" and "system failure behavior".In order to reduce the sampling overhead of the observed data,the LSTM network model in the Recurrent Neural Network is used to predict the timing values of the observed nodes,and the Bayesian network is cascaded to predict the system failure type.By adjusting the number of sampling times in the test,the optimal trade-off between the sampling number and the prediction accuracy of the cascaded model is determined,and the number of sampling times is reduced from 50 times to 39 times which reduces the data collection time by 20% when the original prediction accuracy of the cascaded model is maintained.