Dynamic Safety Margin And Uncertainty Assessment Based On RVM-PF

Quantification of Margin and Uncertainty (QMU) is a new safety assessmenttechnology which mainly base on the concept of “Margin Design”. It pays attention tointerval analysis and considers the actual range of system performance parameters.However, QMU is indeed a static assessment method. Generally, it takes the systemstatic data to support the assessment, which can’t handle the dynamic situations. Inaddition, QMU only provide the assessment method, the data analysis methods ofsystem performance data are not included. And this part is the input of QMU, which hasdecisive influence on the assessment results. For operating system, obtaining thereal-time safety state of system is very important. If the time-series of systemperformance data can be provided by conditional monitoring and prognosticstechnology, the real-time QMU assessment and the prediction of system safety state canbe obtained by putting these time-series data into QMU assessment model, and this is animportant reinforcement of QMU. Given the real-time QMU assessment result, theprotected actions of accidents could be taken in advance, which can reduce thepossibility of system faults or lighten the damage of accidents. This will bring positiveeffects on system safety as well as availability.Using this idea, Relevance Vector Machine (RVM) and Particle Filer (PF) modelare used to analyze system performance data and predict its trend. Chapter twodescribes the basic principles of classic RVM, as well as an improvement on noise partof classic RVM, which uses variance revised hyperparameter to revise the variance inmodel and lets RVM have better adaptability of noise. The regression performanceanalysis of RVM model are taken, which discuss the influence of kernel function’sbasiswidth and revised noise, also the process of building RVM regression database isproposed.Chapter three puts forward the RVM based Particle Filter Prognostic Model, whichusing RVM regression database to figure out particle sampling suggested distribution,particle weight and moving function. It can satisfy the data requirement of Particle Filtermodel. Then the assessment of prognostics is taken, which analyzes the influencebrought by predict length, smooth coefficient and weight updating rate, also acomparison between this model and normal curve fitting predict model is made.Motivated to improve the shortage of traditional QMU (post-analysis and judgingsafety state only by whether confidence ratio is above1or not), chapter four proposes adynamic safety margin and uncertainty assessment model, which can assess the safetystate during the system operating process using the real-time monitoring data andprovide confidence ratio prediction. This improved QMU model refines the assessment rules, which solves the problem of nonmonotonic degradation in dynamic system. Alsoit can control the misjudge rate of “abandon true” and “accept false” to adapt thedifferent safety assessment credibility requirement.Finally, a case study is undertaken to make a dynamic QMU assessment ofrefrigerant pipe of marine reactor (austenitic stainless steel).