Study On The Method Of Fault Diagnosis And Its Application Based On Manifold Learning And One-class Classification

With the machinery and equipment become large-scale and automating, the equipmentcomponents become more and more closely. Once a component is failure, it is easy to cause a chainreaction, resulting in great damage to the equipment and huge economic losses. The structure ofmodern machinery and equipment are complex, influence factors are changeable and the behaviorsare polymorphism,so it is often hard to accurately diagnose the causes of failure and asses theoperational status of the equipment.It has an important significance to timely and effectively diagnosisthe cause and location, analyze the law and cause of faults to prevent or avoid the significant lossesfor fault. So, it’s the key of improving the accuracy of fault detection and diagnosis to extractfault characteristic information from the collection of complex high-dimensional nonlinear data, andanalyze failures according to the characteristic information. In this paper, the fault diagnosis methodbased on manifold learning and Sphere Structure one-class classification is studied, and research intothe fault diagnosis experiments for the rotor and rolling bearing. The main work is as follows:Firstly, the paper studies the research status and the development of machinery fault diagnosismethods from the areas of feature extraction and pattern classification. Analyses the advantages ofManifold Learning which can be used to extract the core data structure feature from thehigh-dimensional nonlinear fault data, and the characteristics of one-class classification algorithmwhich only need one class of samples to realize machine learning. So it has a great significance to usemanifold learning and one-class classification in mechanical fault diagnosis.Secondly, the paper introduces the Laplacian Eigenmaps(LE), which is one of manifold learningmethods. The neighborhood factor k and the choice of embedding dimension d in this algorithm arevery important for extracting the characteristics of low-dimensional manifold. In order to get optimalparameters to make dimension reduction the most efficient close to the original data in the topologicalstructure, a grid search method combined with Silhouette is used in this paper to evaluate the qualityof LE dimensionality reduction and improve the performance of the algorithm.Thirdly, the paper studies the theory of one class classification algorithm, optimizes the relevantparameters through a grid search method combined with cross-validation method. And apply one classclassification to multi-classes recognition., the essence is to identify one kind of samples with a supersphere structure and identify different kinds of samples with many super sphere structures.Fourthly, the method based on the Laplacian Eigenmaps (LE) and sphere structure one classclassification is used in rolling bearing fault diagnosis and rotor-stator radial Rubbing fault locationidentification. The optimized Laplacian Eigenmaps is used to extract the sensitive features of the faultsamples, and then, the characteristics of the samples were input to the sphere structure one-classsupport vector machine to identify different fault samples. The results show that the method is correctand effective in the field of rolling bearing fault diagnosis and Radial Rubbing fault locationidentification.