Mechanical Fault Feature Extraction Method Based On Stochastic Neighbor Embedding

Feature extraction is the base of machine fault diagnosis. How to extract fault feature effectively is one of the hot research topics in the field of fault diagnosis. With the rapid development of the data mining technology, dimensionality reduction algorithm is applied in the intelligent fault diagnosis for feature extraction. Aiming at the nonlinear, high-dimension and complex characteristics of monitoring data, the mechanical fault feature extraction method based on Stochastic Neighbor Embedding is studied in this work, and the related work is as follows:Euclidean distance did not provide a larger relative distance between high-dimensional data points, and might not express the differences between high-dimensional data points well. This paper proposed an improved Manhattan distance based on Stochastic Neighbor Embedding (Manhattan-SNE) algorithm. The algorithm used Manhattan distance to measure the dissimilarities between the high-dimensional data points, and then got the conditional probabilities of the high-dimensional and the low-dimensional space data points. Simulation data analysis and classification recognition demonstrated the effectiveness of the improved algorithm.Stochastic Neighbor Embedding (SNE) is an unsupervised dimensionality reduction method. The labeled information of samples is useful for classification, but it is not effectively used in SNE. An improved semi-supervised Stochastic Neighbor Embedding (ss-SNE) is proposed based on Laplacian Regularized Metric Learning. In ss-SNE, distances between the high-dimensional data points are measured by Laplacian Regularized Metric Learning. The analysis of simulation data verifies the validity of the proposed method.Stochastic Neighbor Embedding (SNE) is a batch method and cannot obtain the mapping function between high dimensional space and low dimensional space, so new data cannot be processed incrementally by algorithm. Aiming at improving the shortage of Stochastic Neighbor Embedding in the batch processing, an incremental SNE algorithm is proposed. The algorithm searches the K nearest neighbors of new sample points in high-dimensional and low-dimensional space. The aim of the embedding is to match the distributions between the two spaces as well as possible. The analysis of simulation data verifies the validity of the proposed method.At last, the above methods are used for fault diagnosis of gear box. Results show that the above methods could effectively improve the fault diagnosis precision and verify the validity and feasibility of the above methods.