Research On Incipient Fault Fusion Diagnosis Methods For Rotating Machinery Based On Manifold Learning

With the rapid development of modern science and technology,there are many large rotating machinery equipment such as wind turbines, aero-engines, industrial gas turbines and others in the fields of industrial manufacturing, aerospace, petrochemical, energy metallurgy and defense industry. Compared with the traditional mechanical equipment, rotating machinery is developed towards large, complex and high speed. While these large rotating machinery and equipment would easily cause the serious influence on production efficiency of enterprises, and then result in huge economic losses as well as irretrievable casualties once a fault happened. The key to save the huge economic losses, casualties and catastrophic accidents from the rotating machinery failures is early fault diagnosis for the rotating machinery, which means timely and accurately identifying the fault in the early embryonic stage, making reasonable arrangements for repair and maintenance strategies, and effectively controlling the development of the failure to ensure the safety and reliability of large rotating machinery.The complex structure of the large rotating machinery, harsh working environments and variable operating conditions result in the strong noise interference and nonlinear characteristics in the vibration signals. The early fault features submerged in strong noise environment are too weak to be extracted. The early faults of large rotating machinery, which are continuously developed, are bred in the normal equipment operation process. The mapping relationship between the fault causes and the results is fuzzy and lack of samples. The state information acquired by a single sensor can only reflect the local operation state of the rotating machinery. Large rotating machinery equipment needs to obtain vibration data through multiple sensors to characterize its overall operating state. Need to solve the problem of decision fusion for multi-source and multi-model. Decision fusion is needed for the independent local diagnosis of multi-sensor multi-recognition model.In this paper, the problem on nonlinear feature extraction, small sample fuzzy information fault identification, multi-source and multi-model decision fusion has been studied. The research content is as follows based on manifold learning method:Aimed at the difficulty of extraction of nonlinear subtle feature, fault recognition, multi-source multi-model decision fusion diagnosis and other issues under the condition of noise interference, this paper has well studied the fusion diagnosis of rotating machines’ early fault based on manifold learning. The detailed information is listed as below: for the problem on extraction of nonlinear weak feature of the large rotating machines’ vibrating status, weak feature fusion extraction of the large rotating machines based on adaptive manifold learning is put forward.The combination of adaptive local tangent space arranging manifold learning and phase space reconstruction is adopted to nonlinearly de-noise the strong-noise vibration signals of rotating machines. After the construction of hybrid domain high-dimension feature vector, the whole running status of the rotating machinery is depicted thoroughly, completely and deeply. Multi-criteria fusion evaluation sequence based on Dezert- Smarandache Theory is proposed to execute the feature selection and eliminate the interference features in original high-dimension feature vector. Finally we will eliminate the redundancy feature in feature vector and achieve fusion extraction of nonlinear weak feature.Aiming at the problem caused by the lack of fault samples and vague projection relations between the fault reason and fault character in the fault diagnosis of large rotating machinery, a supervised fuzzy C mean clustering method utilizing the manifold distance metrics is proposed for the rotating machinery fault detection.The supervised fuzzy C mean clustering extending the value range of membership degree on the basis of hard clustering, has a better description of the fuzzy information and is able to put the prior empirical information in the unsupervised clustering process. Finally realize the improvement of the clustering performance; as for the small-sample circumstances, this algorithm preserves the fuzzy classification of supervised samples and can take advantage of the typical supervised samples by using supervised proportionality coefficient to guide the clustering; once error occurs for the supervised samples, it can effectively reduce the influence of noisy supervised samples on the overall classification performance; fusing the label information of supervised samples by using the non-parameter kernel density estimation, setting the initial clustering center, improving the iteration speed can be helpful; using the manifold distance metric to process the real data which have a complex space distribution makes the algorithm more common for people to realize the effective detection of early fault from the vague information of rotating machinery in the circumstance of small samples.In order to solve the limitation of single sensor vibration information fault diagnosis for the large rotating machinery, a multi-source and multi-model weighted decision fusion method based on fuzzy consistency matrix have been proposed in this paper.By setting the threshold of the conflict coefficient, we legitimately come up with Dempster- Shafer theory or Dezert- Smarandache theory to fuse multi-source and multi-model decision.The reliability of each sensor is reflected by the pattern recognition accuracy of the training sample.The weight coefficient of the multi-source and multi-model are calculated by the fuzzy consistency matrix. The sensor fault recognition results are weighted to increase the tolerance and robustness of decision fusion diagnosisEffective fusion of multi-source and multi-model local diagnosis result for rotating machineryA condition monitoring and early fault diagnosis system for large rotating machinery with self-property right is developed successfully in this thesis. It realizes remote data acquisition via network, signal analysis, early fault feature extraction, fault diagnosis and decision fusion. At last, all the functions of the system were tested though experiments and applications.At the end of the thesis, the work of this paper is summarized, and expectation of the relative technology development is presented.