Fault Diagnosis For Turbine Vibration Based On Manifold Learning

As the core of the plant equipment, stable and secure operation of the turbine-generator unit plays an important role in the national economy and life. Steam turbine generator is complex, and the failure will seriously affect the operation of the plant. The study of turbine-generator unit fault diagnosis technology has important practical significance.For fault diagnosis of steam turbine rotor vibration has the characteristics of the high amount of dimensionality and nonlinearity which is difficult to effectively deal with, this paper researches the nonlinear dimensionality reduction method----manifold learning theory and methods. Firstly, this paper has a comparative analysis of the manifold learning and PCA feature reduction method. And this ensures the applicability of manifold learning theory in the feature extraction for turbine rotor vibration fault. Secondly, this paper builds a mathematical model of turbine vibration fault feature extraction based on manifold learning. Finally, as the manifold learning algorithms embedded dimension, the neighborhood size and other parameters are difficult to be given in advance and this results that the feature extraction is difficult to be done, this paper researches the guidelines of the determination of the optimal parameters. This feature extraction model is applied to the turbine rotor vibration signal of a typical fault feature extraction process. Then feature extraction samples are input the neural network of the fault diagnosis and the test samples are diagnosed. The results show that manifold learning feature extraction in the case of maintaining the original information, filters out the irrelevant variable information, improves the classification performance of the fault samples, greatly reduces the computation of neural network and improves the effective and accuracy of the diagnosis.The feature extraction theory of manifold overcomes the high dimension and non-linear of steam turbine fault samples, preserves the original characteristics of the non-linear information, improves the speed of diagnosis and provides a new way to solve the fault diagnosis feature extraction.