Research On Fault Diagnosis Of Hydropower Unit Based On New Novelty Detection And Support Vector Machine

Research and development of hydropower unit condition monitoring and fault diagnosis technology is a hot topic in hydropower industry recently. On basis of unit operating state data displayed and accumulated by condition monitoring system, and evaluation of unit operating status by data analysis and fault diagnosis, we can find out the trend of unit potential fault and performance deterioration, finally, try to fulfill the unit condition maintenance. Now lots of condition monitoring systems are installed in hydropower plants, and then, it is a bottleneck problem that how to use these huge numbers of long term stored data to push forward the further application of condition monitoring system, and carry out the research of diagnosis methods.In the process of analysis and fault diagnosis for hydropower unit, the problem just like the lack of prior knowledge, rare faults sample and imcomplete fault mode often exist. Most fault diagnosis technologies for hydropower unit are based on knowledge or rules, which have significant effects for specific faults condition data, but as low rotating speed equipments, hydropower units can hardly get clear fault mode and characteristic data corresponding to those mode for the complext operation mechanism. Aiming at the above problems, this paper takes long term data of hydropower unit vibration, displacement, temperature and other common condition measurement points as the research objects, makes research about hydropower unit diagnosis on basis of new novelty detection methods and support vector machine.The main writing idea of the paper starts from the analysis of symptom statistics characteristics, to modeling of new novelty detection methods (single class support vector machine) and research of evaluation methods, then extends to multi kinds of fault diagnosis and trend evaluation based on support vector machine. The first part of the paper analyzes fault diagnosis basic conditions—characteristics extracting method of fault symptom, choosing condition data long-term statistics characteristics as model study sample, melt new heterogeneous detection (single kind support vector machine),and realizing abnormal evaluation of hydropower unit state data. Second half of the paper spreads from single kind support vector machine to multi kinds of fault diagnosis classification, and finally, combining units state monitoring dada characteristics, designs and promotes two methods of support vector machine regression and online data update trend evaluation.In the paper, research work is mainly as follows:1. To analyze the significance of new heterogeneous detection method to hydropower units, and taking mode recognition and fault diagnosis method classification in machine study and other energy fields as reference, summarize hydropower units fault diagnosis mode, and promote that hydropower units fault diagnosis methods can be divided into knowledge-based diagnosis method and data-based two categories.At the same time, through the summary of new heterogeneous detection technology research and application status, new heterogeneous detection technology and single kind support vector machine which roots in support vector machine are thought to be possible to judge and recognize unknown abnormal states according to observational study of known samples.2. Statistics characteristics of hydropower unit state data are intensively studied. By analyzing time domain statistics characteristics of hydropower unit state data, design healthy sample model basing normal operating data, and have made mathematical analysis and verification of the long-term distribution characteristics, short-term distribution characteristics and time sequence characteristics. Finally, Bootstrap sample method is applied to the statistics characteristics extraction of state data, and promotes a threshold value calculation method of hydropower units state characteristics based on Bootstrap method, and evaluates its effect by examples.3. Use statistics characteristics and single kind support vector machine technology to realize unit state signal melt detection.(1) Introduce unit state evaluation method basing on experience data in details, verify this method combining unit actual state data, and promote the choosing basis of health model reference value.(2) Combining unit statistics characteristics and Gaussian density method, a three dimension health model built method basing on unit normal operating data is designed. By operating condition grid division, remove the effect on state data from operating condition, and optimize health model threshold value scope.(3) Promote a unit state signal detection and evaluation method basing on new heterogeneous detection method. By melting unsupervised clustering and single kind support vector machine method, unit state trend signal abnormal detection and recognition are realized.4. Promote a hydropower fault diagnosis method based on wavelet packet energy decomposition and least square support vector machine.(1) Introduce multi kinds classification method and overall situation optimization program of support vector machine technology in details. By wavelet packet energy decomposition technology, taking unit shafting characteristics data as example, through multi frequency band energy decomposition, shafting fault sign extraction is realized.(2) Combine wavelet packet decomposition technology and least square support vector machine, design diagnosis method procedure from sign extraction to sample study and then to fault classification, and verify the results by unit actual operating data.5. Basing on online forecast method on basis of support vector machine regression technology, promote sectional forecast and multi-scale forecast method to hydropower unit trend forecast of real time updated state monitoring data.(1) Aiming at calculation overload, low efficiency problems caused by normal state monitoring system data update, promote support vector machine regression method basing on sectional forecast. By choosing best sub classification support vector machine model to forecast output, realize online sample scale reduction, save sample historic characteristics, and improve calculation method generalization ability. The method is proved by tests that it can maintain high forecast efficiency, and improve forecast accuracy of online monitoring data sample.(2) Aiming at the problem that state monitoring system data update is easy to lose trend characteristics, promote a parallel forecast method basing on multi-scale analysis. By multi-scale data re-construction to sample sequence, realize sample scale reduction, use multi support vector machine regression model to parallel forecast, choose best model to forecast output, and improve forecast accuracy significantly.Above all, avoiding the features that normal hydropower unit fault diagnosis technology research is usually based on fault prior knowledge, the paper makes research on unit trend data statistics characteristics and normal data modeling. Using new heterogeneous detection method and support vector machine technology, melting with clustering calculation method and wavelet packet energy decomposition technology, and aiming at online monitoring data update features, it promotes the improvement and optimization of support vector machine, and realizes the unit state data fault classification and online trend forecast.