Research On Dimension Reduction Of Rotor Fault Data Based On Manifold Learning

In recent years, with the rapid development of information technology, how will information technology into rotating machinery and equipment, and use based on data driven theory to excavate an intrinsic property of real-time reflect the mechanical operation method has become the inevitable trend of the development of mechanical fault diagnosis. Manifold learning algorithm as a new construction of spatial data structure in high dimensional nonlinear data processing algorithm, which has been successfully applied in many fields, such as: data mining, data pattern recognition and classification, fault diagnosis and so on. However, isometric mapping(isometric mapping, Isomap) and local linear embedding algorithm(locally linear embedding LLE) by the two more classical manifold learning algorithm shows their respective advantages and disadvantages in the process of data processing. Therefore, this paper mainly combines the advantages and disadvantages of these two algorithms in the theory of kernel function, and then puts forward a new KISOMAPLLE algorithm which can take into account the characteristics of these two algorithms. The main research results are as follows:(1) A new KISOMAPLLE algorithm is proposed by fusing the two algorithms in the kernel framework, and it is successfully applied to the six artificial data sets. The experimental results show that the proposed algorithm can achieve dimensionality reduction successfully in the artificial data set, and it has the characteristics of weak sensitivity and strong robustness to the neighborhood.(2) Will KISOMAPLLE algorithm and other 7 kinds of algorithm is applied in original rotor fault feature set, and data dimensionality reduction and comparative analysis. The experimental results show that the fault mode identification algorithm is proposed in this phase compared with other algorithm has higher accuracy.(3) Of the algorithm and several other algorithms in rotor fault data concentration varies with neighborhood relations were related research, experimental results show that this algorithm phase compared with several other algorithms of neighborhood selection is less sensitive and in synthetic data experiment results have the same conclusion, and then validate the proposed algorithm to solve the practical problem has certain validity and feasibility.(4) Will support vector machine(SVM) were optimized by particle swarm optimization algorithm, classifier, the experimental results of experiments and research, and on 8 kinds of algorithms of dimensionality reduction results are applied SVM and PSO-SVM the accuracy analysis and comparison. Experimental results show that the classifier of fault recognition rate has a certain degree of influence, but for the algorithm was less affected, so that the algorithm proposed in this paper has the classifier selection modific weak, generalization ability is strong and so on characteristics.