Research On Kernel Based Fault Recognition And Condition Forecasting For Mechanical Power And Transmission Systems

The large-scale mechanical power and transmission systems are the types of important technical equipment being used extensively in national defense and national economy realms; their safety and reliability are crucial. Because of a series of characteristics including complicated structures, serious states coupling among the different parts, a strong nonlinear on their condition behaviors, an abominable operational environments and being subject to the non-gauss noise and various indetermination factors, it is quite difficult to carry out the missions of the fault diagnosis accurately for these systems. As an important supporting technology of ensuring that the machines operate safely and reliably and their maintenance actives are organized by the scientific and reasonable manners, researches on the advanced techniques including condition monitoring, fault diagnosis and condition forecasting are of quite significance.Based on analyzing the nonlinear characteristics and aiming at the technical requirements of nonlinear fault diagnosis and also combining the recently achievements deriving from related academic realms, the kernel-based diagnostic techniques including feature extraction, fault detection/recognition and decision-making, the operational condition forecasting for mechanical power and transmission systems are studied systemically and detailed. The main purpose of the research work is to investigate the new approaches to deal with the problems of fault identification and condition forecasting accurately for the systems. At the same time, the research work also aims at providing valid technical support for promoting the development and application of the fault diagnostic technology. The detailed contents and innovative works are shown in the following sections.1. To meet the requirement of extracting the effective features for the fault diagnosis of the mechanical power and transmission systems, the kernel-based methods of nonlinear feature extraction are studied. Firstly, the method of the debris feature extraction based on KPCA is presented. It is mainly used to analyze the nonlinear related features and reduce the dimension of the debris parameters. Secondly, the preprocessing technique for the fault feature extraction based on KICA being used in the helicopter transmission system is put forward. The research result demonstrates that the KICA is an effective approach of extracting the features that reflect the incipient faults. Thirdly, the KFDA based method of fault feature extraction is presented. It is good at enhancing the performance of classification for the fault features and promoting the veracity of the fault recognition.2. To solve the diagnostic problem based on a small quantity of fault sample existed in large-scale mechanical power and transmission systems, the SVM based method of the fault recognition and decision-making is studied in detail. To promote the performances of two schemes of multi-classifier combination including 'one against one' and 'one against rest', the improved methods are proposed respectively, which is good at the training efficiency and can keep the original precision in classification. Considering that the different mistakes of decision-making may lead to unequal loss and risk in fault diagnosis process for the mechanical power and transmission system, the equal risk SVM model based on fuzzy membership function for fault diagnosis is presented. To meet the requirements of online diagnosis for mechanical system, the SVM based multi-layer model for fault detection is studied. Its validity is confirmed by fault detection experiment on bearing fault detection and classification.3. For the purpose of fault detection for the large-scale mechanical power and transmission systems under the condition of the lack of fault samples, a novel fault detection method based on the SVDD model is put forward. The influences on decision-making boundary and classification precision by choosing different kernel functions and the kernel parameters are deeply studied and analyzed. An improved SVDD based model for fault detection is presented and discussed, which considered unequal loss owing to two kinds of wrong decision-making. The presented method is used for fault detection of the bearings mounted in a helicopter's transmission system and its validity is proved by the experimental studies.4. Aiming at nonlinear characteristics existed in running condition of the mechanical power devices and combining the support vector regression with phase space reconstruction theory, a novel condition forecasting method is put forward. It is mainly by means of analyzing the nonlinear time series depicted the running condition changes and calculating the minimize embedded dimension which is used for determining the dimension of the sample's vectors by using the phase space reconstruction technology. The model of support vector regression is employed to forecast the running condition and its change trend. The forecasting capability of the model is investigated in detail. The related techniques have been applied successfully in the running condition forecasting of the mechanical power device in a warship. The research results indicate that the presented methods can provide a valid solution for the large scale and complicated machines' condition forecasting.