| Against the background of green energy resources in the new era, it is of great strategic significance to vigorously develop the hydropower energy for the energy structural optimization and the economic sustainable development in our country. As the core device of energy conversion in the hydropower plant, the development tendency of hydropower units possesses giant sized, high-speed, supercritical and intelligent features, which results in growing fault security problems under its running conditions. For the sake of reducing the fault accidental risks and ensuring safe and steady operations of the units, a consensus of actively carrying out researches on fault diagnosis of hydropower unit has been reached in engineering practice and academic field. In practical applications, apart from the complicated structural characteristics of the unit and multiple excitation source interferences, the shortcoming and limitation in the existing fault diagnosis theories prevent the fault diagnosis theories from transforming to engineering applications.To this end, addressing on the topic of non-stationary signal analysis and feature extraction, multidimensional features reduction and complex fault pattern recognition in hydropower unit fault diagnosis, this paper has investigated and overcome the theoretical and practical defects in the existing methods and has proposed several effective modified schemes in non-stationary signal analysis and feature extraction specific to hydropower units by taking empirical mode decomposition (EMD), manifold learning, hybrid gravitational searching, support vector machine (SVM) and random forests as the theoretical basis and study means. Focused on the multi-dimensional feature redundancy interference and the like, two feature reduction methods based on space mapping and intelligence searching have been built, respectively. Finally, the time-frequency analysis and features reduction incorporated fault diagnosis model for hydropower units is established of certain practical means by introducing SVM and random forests fault recognition theories, which provides theoretical novelties and new research ideas to study in the very field. The main research contents and novelty in the paper are listed as follows.(1) In reaction to the ends swing phenomenon, the root cause and its solution have been extensively studied. An end effect elimination scheme for extreme value point mixed extension-based EMD, which integrates the privileges of support vector regression and mirror couple extension, has been presented. At the preliminary stage for signal decomposition, the extreme value point sequences of the signals are extended on both sides with the utility of support vector regression for its powerful prediction capability at near-end. Next, for the low-frequency component, problems of lacking training sample data at the extreme point of support vector machine regression and poor inhibition effect are avoided by mirror couple extension method. Eventually, multi-fault vibration fault simulation signals of hydropower units are decomposed by the proposed method. Compared with the mirror extension and the support vector machine regression signal sequence extension methods, the results indicate the effectiveness in end effect suppression and the lower time consumption in computation of the proposed method.(2) Taking account of non-stationary and noise impact characteristic, the ensemble empirical mode decomposition (EEMD) time-frequency analytical approach is introduced in order to enhance the suppression to mode mixing in EMD. In reaction to the existed hardships of EEMD in parameter tuning, calculation burden and residual noise pollution, an adaptively fast EEMD algorithm is proposed intuitively on the basis of complementary EEMD. The study illustrated the eliminating mechanism of mode mixing under minor noise amplitude condition in EEMD and set the magnitude of white noise and the number of ensemble trials to be 0.01 times to the standard deviation of target signal sequence and 2, respectively, based on the aforementioned reasons. A relative root mean square error index-based adaptive determination method for the upper limit of white noise has been proposed in order to breakthrough the conundrum of choosing the key parameters of EEMD. The validity and the advancement of the proposed method has been verified by both simulative tests and international standard cases for fault diagnosis in hydropower units. At last, the proposed method is applied to the analysis of pressure fluctuation in draft tube of a hydro turbine, in which the engineering practicality is testified.(3) For the sake of attenuating the feature redundancy, the sensitive information submergence and the like existed in high dimensional feature space, we have built the measuring index of Pearson correlation coefficient in neighborhood relationship of sample points and then proposed a novel data reduction technique for adaptive neighborhood-related supervised local and global principal components analysis. Meanwhile, considering the small sample property of hydroelectric generator unit faults, a popular learning feature reduction and SVM incorporated multi-fault diagnosis model is established. Its validity has been testified by real fault cases of the power plant and contrastively analyzed with other feature dimension reduction methods. In addition, for fault diagnosis problems of greater complexity, a multi-dimensional generalized featured vibration fault hybrid diagnostic strategy is proposed, which effectively merges probability statistical analysis and data-driven fault diagnosis together and simplifies the diagnostic process into three procedures, i.e., fault preliminary detection, fault type recognition and fault degree determination. Finally, this hybrid diagnosis strategy is applied to fault diagnosis of rotating machinery. The results indicate that the proposed scheme not only enhances the precision of fault diagnosis, but also reduces the complexity of the diagnosis model, thus improving the calculation efficiency and, simultaneously, providing the complex fault diagnosis problem in hydropower unit with an effective alternative.(4) Focused on the problems which are difficult to evaluate the quality rating of every feature with space mapping-based feature reduction methods and for the sake of eliminating the negative effect on the random forests diagnostic precision caused by model parameters, this chapter proposes a hybrid gravitational searching algorithm (HGSA) and random forests based fault diagnosis model from another perspective of feature reduction. The model have made optimization and modification to the random forests fault diagnosis from feature subset and model parameter aspects. Binary GSA is utilized in searching for the optimal subset for the choice of the feature subset, meanwhile, real-valued GSA is applied to optimize the number of the decision tree of the random forests. In the end, the proposed method is applied to the rotor system fault diagnosis. According to eight diagnosis experiments of different fault types, it has been proved that the proposed method enables to get rid of the interference of the redundant feature information and solve the difficulty in choosing the number of decision trees in random forests as well as improve the precision and efficiency of the random forests diagnosis model. Furthermore, the importance evaluation of the optimal feature attributes has been achieved by using the estimation of out of bag data in the random forests. |
PDF Full Text Request