Research On Life Prognosis Methods Of Nuclear Power Plant Assets Based On Condition Monitoring Data

Important assets are typically monitored by various online or offline condition monitoring measures in nuclear power plants,producing massive condition monitoring data.Information related to the performance status or degradation omen of monitored assets is usually contained in these data.This dissertation is devoted to mining this information by establishing life prognosis methods for important assests based on condition monitoring data,in order to accurately predict performance degradation trend and residual useful life of the corresponding assests.This will provide technical basis for proactive aging management and maintainence management,which may be the next step towards the ultimate goal of ulteriorly improving the safety and economic level of nuclear power plants.Different prognosis models are established for active and passive assets according to the different characteristics of their prognosis problem.Ensemble data driven prognosis models are adopted for active assets represented by rotating machinery.Firstly,a “fuzzy clustering dynamic ensemble” method is proposed.The whole training set is devided into subsets by fuzzy clustering method,based on which sub-models for life prognosis are trained.The clustering process improves the diversity of sub-models,which is helpful for achieving a good ensemble.Weights of the sub-models are allocated dynamicly in the prediction process based on similarity between the test data and the training subsets of the corresponding sub-models.Verifications by shared benchmark data sets show that the proposed method can achieve higher prediction accuracy by the cost of less training cost,and the prediction performace of the propsed model ranks in the top level among published related works.Additionally,heterogeneous ensemble model is utilized to further improve the performance of the prognosis model.Two new model integration strategies,i.e.secondary learning machine model and segmented weight allocation model,are developed as extensions to the conventional fixed weight integration strategies.Verification by benchmark data set manifests the superiority of the proposed strategies compared to conventional ones.On the other hand,hybrid prognosis models are adopted for passive assets to integrate information from degration mechanism and reference data.Specifically,in cases where the degration mechanism is unavailable,surrogate type hybrid models are adopted.Uncertainty analysis of the data driven surrogate model is an important problem in the development of surrogate type hybrid models.Guassian process regression model is studied to achieve more convenient model uncertainty assessment,and is verified by simulation data.Further,in order to improve the adaptability of surrogate models,adaptive parameter data driven model is proposed,in which model parameters can be adaptive adjusted by prediction error in the online prediction process,thus making the surrogate model more consistent with monitored data.The effectiveness of adaptive parameter data driven model in improving model adaptability and prediction accuracy is verified by a numerical example of performance prediction of a reactor system.Finally,in cases where mechanism model is available but affected by large bias,modification type hybrid models are adopted.A non-parametric data driven modification model based on Gaussian process regression is proposed,where the original mechanism is modified through a scale factor,an offset factor and the Gaussian process term.Simulation data of flow accerelated corrosion is used for verification of the proposed modification model.