Hybrid Intelligent Fault Diagnosis Method Integrating Support Vector Machines For Hydroelectric Generator Units

With the rapid development of hydropower resources, the proportion of hydropower energy in electrical energy structure is gradually increasing, and hydroelectric generator units (HGU) which are the key equipment of the hydropower production process are becoming more and more large-scale, complex, high-speed and high-power. Meanwhile, the integrated degree of the HGU is becoming higher, and different parts of HGU influence each other, which makes the dynamic behavior of HGU become more complex. Consenquently, the vibaration problem of the HGU has become increasingly prominent, and the vibrant influence on the safety and stability of the grid has become increasingly large. Hence, the conventional vibration fault diagnosis methods for HGU are not adapted to current engineering practice. There is an urgent need to adopt some effective intelligent fault diagnosis methods to improve the accuracy, intelligence and robustness of fault diagnosis for HGU. To solve the Key scientific issues in engineering applications of fault diagnosis for HGU, support vector machine (SVM) is used to diagnose the vibrant faults. The theory and engineering application of SVM are deeply researched. To compensate for the shortage of SVM in engineering practice and bring the performance of SVM into full play, advanced signal processing techniques and some other intelligent methods and SVM are fused to form some hybrid intelligent fault diagnosis methods. The main contents and innovative achievements of the paper are as follows:(1) Model parameters selection for SVM and the influence it affected on the SVM performance are researched. Several forms of inter-class distance in the feature space (ICDF) are introduced and discussed. And, the ICDF is selected as one measure heuristic to shorten the range of the kernel parameter. Then, self-adaptive differential evolution algorithm is used to search the optimal parameter combination in the continuous intervals of kernel parameter and penalty parameter. Engineering results show that the proposed method can effectively locate the typical failure of the HGU. Thus, the proposed method is feasible and effective in fault diagnosis for HGU. (2) Ensemble empirical mode decomposition (EEMD) and Hilbert spectrum and Hilbert marginal spectrum are used to analyze the draft tube pressure fluctuation. Then, the research is emphasized on feature extraction for fault vibrant signals. Two types of features, the EEMD energy entropy and singular values of the matrix whose rows are intrinsic mode functions (IMFs), are extracted. EEMD energy entropy is used to specify whether the machine has faults or not. If the machine has faults, singular values are input to multi-class SVM optimized by inter-cluster distance in the feature space (ICDSVM) to specify the fault type. The proposed method is applied to rotating machinery fault diagnosis. The results show that the proposed method can identify a variety of operating conditions. And, the proposed method has been successfully applied in the fault diagnosis system of Songjianghe power plants.(3) Fuzzy SVM is used to diagnose faults of HGU. In the training phase, by giving different weights for different training samples, fuzzy SVM can effectively eliminate the influence of outliers or noise points for fault diagnosis. Fuzzy sigmoid kernel function is used in the fuzzy SVM, and the advantage of this fuzzy sigmoid kernel function is also described. To solve the problem that the membership function (MF) of fuzzy SVM is difficult to determine in applications, a new method based on reverse k-nearest neighbor algorithm (RkNN) and euclidean distance between class means is proposed to determine the MF for fuzzy SVM. On the basis of deep analysis of one-against-one multi-class SVM, we point out that not all SVM training is necessary for the final decision surface. To remove unnecessary SVM training, one form of improved one-against-one multi-class SVM is proposed. The proposed method is used in fault diagnosis for HGU, and the results are satisfactory.(4) Traditional fault diagnosis classifier can not reflect the uncertain information in fault pattern recognition. To overcome this disadvantage, a novel classifier based on rough set and multi-class SVM is proposed. The proposed method takes full account of the advantages and disadvantages of SVM and rough set theory, and it is an organic integration of these two methods. In the proposed method, the essential ideas of rough set: upper approximation and lower approximation are used to describe the classification results of SVM, which can well describe the uncertain information in fault diagnosis for HGU. Hence, the proposed method takes full advantage of the strong modeling capabilities for uncertain data of rough set and powerful generalization ability of SVM. Engeering application shows that the proposed method can well diagnose coupled faults and Sub-health state of HGU. We also conduct a comprehensive analysis on the "too large swing" problem in upper guide bearing of # 3 unit of the Ertan Hydropower Station. The analysis conclusions provide useful guidance for the manager and operator of the Ertan Hydropower Station, also further illustrate the need of fault diagnosis for coupling faults of HGU.