Researches On Fault Diagnosis For Reactor Coolant Pump Of Nuclear Power Plant Based On Vibration Analysis

China has been becoming the second big country of energy consuming. As thecontinual growth of economy, it is the energy problem that has been the importantreason of constraining the economy development. In such a problem, power also hasthe most important role. Among all kinds of power, the nuclear power, as a new,cleaning energy, provides benefits for our shortage of power.Nuclear power comes from nuclear power plant. The importance and sensitivityof nuclear power plant has got more concerns. Especially after the leak accident ofnuclear power plant in Japan, the safety of nuclear power plant became hot issuesonce again. The safety has been a problem that must be resolved in the developmentof nuclear power. If the safety of nuclear power’s generation could not be guaranteed,not to mention the development that will go on. It is only the safe generation ofnuclear power that can make the new energy sustainable and the economy and societybenefit from it.Among all the devices in nuclear power plant, the core machinery equipment isreactor coolant pump in nuclear island. It also plays an important role in nuclearpower plant, because it can make the heat exchange by driving the coolant flowcircularly in a loop. Whether the reactor coolant pump runs normally or not directlyaffects the reliability of all the nuclear power devices and the safety of nuclear powerplant. The research object of this thesis is the vibration fault of reactor coolant pumpin the primary loop in nuclear power plant. By using the hardware data acquisitionplatform we made by ourselves, some vibration signals are acquired when reactorcoolant pump is running. According to these vibration signals, a fault diagnosis modelis built. The thesis focuses on the fault diagnosis of reactor coolant pump, studyingfurther on data acquisition, feature extraction, fault classification and intelligent fault diagnosis. The main researches and conclusions include:(1) A completed FDSRCP system is proposed in this thesis, based on thecharacters of reactor coolant pump, building the hardware data acquisition platform,feature extraction and fault classification. Currently, the study of the fault diagnosis ofreactor coolant pump in China is still in its initial stage, a complete architecturesystem is not designed. This thesis gave a positive research on the fault diagnosis forreactor coolant pump in nuclear power plant. It provided technical support for thedevelopment of fault diagnosis for reactor coolant pump.(2) We developed a high precision data acquisition module PXI2120byourselves based on the general bus platform of PXI instrument-oriented. PXI2120issuch a new data acquisition module that uses16-bits high resolution A/D, combingthe technologies of FPGA with PXI. This platform can acquire the vibration signalswhen reactor coolant pump is running. These collected data will be sent to maincomputer through the PXI interface. All the designs of PXI2120are independentlydeveloped, including the design and implementation of sample module, timebasemodule, trigger module, stored module and PXI interface module. PXI2120breaks thetechnological monopoly of foreign companies in the domain.(3) Feature extraction from original data is the key step of fault diagnosis. In thisthesis, vibration severity, wavelet energy spectrum and power spectrum of waveletenergy maximum decomposition level are firstly used as feature parameters in thefault diagnosis system for reactor coolant pump, based on the comparison and analysisof related algorithms of feature extraction. We proposed a method of featureextraction named VSWEPS which is a method of time-frequency. Vibration severity issuitable for feature extraction in time-domain. Vibration severity that can get from thevalue of speed is able to affect how much the vibration energy is. That is why thevibration severity is as a parameter of feature extraction in time-domain. Frequencyanalysis is one of the most important tools in feature extraction, because the vibrationsignals from running devices are non-stationary signal, so traditional Fouriertransform can not be used here. Though wavelet transform can be used to analyzenon-stationary signal, it has no ability to accurate analyze faint signal in vibrationsignals. The faint signal is so important that it is often as a sign of fault. For thisreason, the methods of time-wavelet energy spectrum and power spectrum of waveletenergy maximum decomposition level are employed as a method of feature extractionfor vibration signals based on wavelet transform in this thesis. The experimentsshowed that better effects can be obtained by using VSWEPS as feature extraction for vibration signals, especially at the early stage that a fault would be happened.(4) The classification for fault of reactor coolant pump faces to a problem thatthe sample data is constrained by the characteristic of device. The collecting of datasetis just small sample dataset. The support vector machine (SVM) algorithm based onstatistical learning theory is suitable for training and learning of small sample dataset.SVM uses kernel function to build a mapping relation from nonlinear space to linearspace, which gives a better solution to the problem of dimension disaster. The SVMalgorithms are comprehensively further studied in this thesis. The basic ideas andimplemental algorithms of SVM are discussed in detail. The SVM has beenintroduced into the fault diagnosis of reactor coolant pump. A multi-faultsclassification model based on C-SVM is also generated by recognizing the categoriesof feature parameters. This model can distinguish the categories of many vibrationfaults in one time. The experiments proved the effectiveness and practicality of theSVM-based method of classification.