Research On Model-driven Hybrid Fault Diagnosis For Nuclear Power Plant

Nuclear power plants have been extremely useful for socio-economic development of many nations despite the potential devastating effect it has on the environment if uncontrollable accident occurs.One of the most prominent causes for accidents is human operator error.Recently,effort has been directed towards offshore operation of this plants due to certain advantages such as islanding operation,limited operators and resources requirement.However,fault detection in the digital control system and related facilities in commercial nuclear power plants have their own limitation,placing heavy reliance on operators.Hence,safe operation of the nuclear power plant is highly influenced by the operator,which does not conform to the safety standard required of commercial plants.Consequently,development of intelligent operator support system is important,to decrease the possibility of operator errors.Additionally,it is also an essential method for intelligent control and emergency support.Currently,one of the most serious problems in operator support technology is the inaccuracy and credibility of their core techniques—— applicable to on-line monitoring and fault diagnosis.Few decades ago,researchers discovered that it is hard to acquire sample data for data-driven techniques under various fault conditions and the results cannot be explained.And for knowledge-based methods,the knowledge is difficult to obtain and the scope of its qualitative reasoning is limited.Also,a single method cannot solve these problems.However,with the development of real-time simulation techniques,the quantitative analytical model provides a possible way to solve the problems.Therefore,the quantitative simulation model is combined with data-driven and knowledge-based methods and they will learn from each other to achieve a hybrid strategy on-line fault diagnosis.Based on this background,this thesis discusses the following aspects:The technical process is improved and the coupling between each technique is defined.Further,the hybrid strategy and corresponding theories are presented.On the whole,quantitative simulation model is utilized as a clue and combined with data-driven and knowledge-based methods.The rationality and validity of them are explained,which will ease the passive situation that leads to distrust of results by operators.Because sample data under abnormalities and malfunctions are rare,on-line mechanism simulation model is built based on thermal-hydraulic processes.On the basis of traditional techniques for simulation modeling,the node partition is improved to not only ensure the detail and accuracy of some significant equipment but also ensure that the model could calculate faster than real-time as required.Furthermore,real-time data from nuclear power plants is introduced to the simulation model which will initialize and modify the running characteristic,auto-control and manual devices.Thus,the simulation models could run in real-time with nuclear power plant.After faults,they could provide sample data faster than real-time for fault diagnosis or trends prediction.By connecting the simulation model with full-scope simulator,the accuracy and efficiency of proposed methods are verified.Since nuclear power plants is a multi-component system,any fault in one of the components is likely to propagate through the system.Therefore,it is difficult to achieve the ideal effect by treating all systems as a whole.By utilizing distributed strategy,the whole plant is divided into several separate units according to the arrangement of sensors.And then,the residual errors of characteristic parameters in each distributed monitoring unit and corresponding real systems are calculated and transferred to the residual analysis module based on Sequential Probability Ratio Test which could detect the abnormalities more rapidly.However,due to sensor limitations,knowledge-based methods have to be utilized after system-level monitoring,they can do causal reasoning to find out the root alarms.In addition,these methods are verified by simulation tests.As the diagnosed results by knowledge-based methods are not always unique and may have some uncertainties,support vector machine(SVM)is utilized for fault type verification and filtering.SVM can be trained using a few sample data and it is not prone to local optima compared with other methods.One of the major challenges with this method is the selection of some hyper parameters.If not properly selected it may lead to a high uncertainty,affecting the accuracy of classification.Thus,this thesis develops an improved particle swarm optimization algorithm by adopting multiple search strategies to maximally ensure the accuracy of fault type verification and also meet the on-line demand.Finally,the accuracy and efficiency of the methodologies are demonstrated through simulation test.For fault severity assessments,because many parameters have a strong non-linear coupling with each other,a small change of some parameters would severely weaken the characteristics of other parameters as all the measurements are utilized.Further,the accuracy of severity assessments cannot be guaranteed.To solve this problem,kernel principle component analysis and other non-linear manifold learning methods are utilized.By comparing the results,kernel principle component analysis is adopted for fault feature extraction.During the fault size estimation,there are a large number of fault types and corresponding many more fault degrees which lead to indeterminate failure modes.In order to solve this problem,a method based on similarity clustering is presented.Several typical distance functions are utilized for clustering after dimension reduction.According to the simulation tests carried out during implementation stage,it is observed that the Euclidean distance function has the best performance.