Software reliability modelling of Reactor Protection System based on change point and severity analysis

Abstract

Nuclear safety has received great concern since Fukushima Accident. As the employment of digital systems in newly-built and upgraded Nuclear Power Plants (NPPs), software reliability brings a lot of challenges to the Probability Risk Assessment of NPPs. Digital Reactor Protection System (RPS) plays an important role in nuclear safety: it monitors the state of the plant and generates an emergency scram signal if certain nuclear accident arises. To meet the requirements of regulation and Probability Risk Assessment, we collected the fault detection data during the RPS software testing process and modeled this testing process with Software Reliability Growth Models (SRGMs). In order to describe the characteristics of the data, we built an SRGM based on non-homogeneous Poisson process (NHPP). Severity analysis was considered by identifying the software faults into easy and hard ones modeled with different SRGMs separately. Goel-Okumoto (GO) model was selected to describe the easy faults and the hard faults were modeled by Inflection S-shaped (ISS) model or Delayed S-shaped (DSS) model. As the fault detection rate and the inflection factor is different because of changes of testing environment, testing strategy, resources, etc., we also used change point (CP) method to improve the fitting and prediction effects. The data collected in our project were used to testify the validity of our models. According to a series of analysis, with CP model has good fitting and prediction abilities even when the data obtained are limited while GO&ISS model has good prediction effect only when there are sufficient testing data. On the other hand, because of the abnormal fitting parameters, model and GO&DSS with CP model are not proper for the fitting of our data. We also find that the adoption of change point can improve the prediction effect on both GO&ISS model and model. The models we proposed give good estimation of the proportion of easy faults and good prediction result, which can be used to guide future testing work.