Online diagnosis of accidental faults for real-time embedded systems using a hidden Markov model

This article proposes an approach for the online analysis of accidental faults for real-time embedded systems using hidden Markov models (HMMs). By introducing reasonable and appropriate abstraction of complex systems, HMMs are used to describe the healthy or faulty states of system's hardware components. They are parametrized to statistically simulate the real system's behavior. As it is not easy to obtain rich accidental fault data from a system, the Baum-Welch algorithm cannot be employed here to train the parameters in HMMs. Inspired by the principles of fault tree analysis and the maximum entropy in Bayesian probability theory, we propose to compute the failure propagation distribution to estimate the parameters in HMMs and to adapt the parameters using a backward algorithm. The parameterized HMMs are then used to online diagnose accidental faults using a vote algorithm integrated with a low-pass filter. We design a specific test bed to analyze the sensitivity, specificity, precision, accuracy and F1-score measures by generating a large amount of test cases. The test results show that the proposed approach is robust, efficient and accurate.