Research On Fault Diagnosis Methods For Vulnerable Components Of Rotating Machinery Based On Semi-Supervised Manifold Learning

The rotating machinery is easy to be damaged in the bad working environment for a long time,which has become a hidden danger to the safety of equipment and personnel.In the actual operating conditions,fault signals often demonstrate non-stationary,non-linear,non-gaussian distribution and other characteristics.Traditional signal processing technology and intelligent diagnosis methods are difficult to extract the accurate fault characteristics.In addition,the acquisition of the label fault samples is time-consuming and laborious.Therefore,the feature extraction method and diagnosis technology based on the semi-supervised manifold learning were deeply studied.The main research contents of this paper are as follows:(1)The process of fault diagnosis based on data-driven is described and the methods of feature extraction and pattern classification of rotating machinery are introduced and compared.(2)A semi-supervised Laplace Eigenmap(SSLE)algorithm is proposed,which has a good feature extraction performance on the complicated high-dimensional signals.Compared with the traditional unsupervised manifold learning method,the algorithm can make full use of the fault information contained in a small number of labeled samples,and get more accurate low-dimensional fault features.(3)A fault identification model based on SSLE algorithm and constrained seed K-means is presented.The model utilized the SSLE algorithm to directly extract the most sensitive low-dimensional manifold features from the raw high-dimensional vibration signals.Subsequently,they were fed into the classifier based on the constraint seed k-means algorithm.Thus,the operating conditions of rotating machinery were identified by the visual clustering results.Compared with kernel principal component analysis and kernel discriminant analysis,the model obviously improves the recognition performance of bearing fault types and ball fault severities.(4)Considering that multi-sensor signal fusion diagnosis can significantly improve the recognition rate of fault,a fault identification model based on SSLE algorithm and deep belief network is proposed by combining with the ideas of manifold learning and deep learning.In this model,manifold learning method is carried out for multiple original high-dimensional Spaces respectively.And then the low-dimensional fault features from multiple sources are fused to perform mining and representation in the deep learning network.Taking the multi-sensor feature fusion diagnosis of gear,the model achieves a good diagnosis effect under the unbalanced training samples,and has engineering application value.