Abstract: Aiming at the problem that the traditional dimensionality reduction method is difficult to maintain the local and global geometric structure characteristics of the datasets, geodesic distance is selected as the measurement index, and an improved t-SNE fault data set dimensionality reduction method is proposed. In this algorithm, firstly, the multi-domain high-dimensional fault data set of denoising vibration signal is extracted and normalized, then D-t-SNE algorithm improved by GD index is used to reduce the dimension of high-dimensional fault data set, remove the redundant information, and then identify the fault modes of the low-dimensional feature subsets through different classifiers. Taking UCI data set and the simulated fault data set of the double-span rotor test bench as experimental objects, the above methods are verified and compared with the implementation results with SNE and t-SNE. The results show that this method can reduce the dimensionality of high-dimensional fault data set, thus it can reduce the difficulty of fault classification and improve the performance of fault identification accuracy. This study can provide a theoretical reference for reducing the scale of the original fault feature data set of rotating machinery, reducing the difficulty of fault classification as well as improving the visualization of fault identification results.

Key words: t-distributed stochastic neighbor embedding, rolling bearing, data dimension reduction, fault diagnosis, geodesic distance