Abstract: In response to the issues of inaccurate neighborhood selection and deficiencies in the metric used in the LLE(Locally Linear Embedding) algorithm, which hinder its ability to extract the true manifold structure, an algorithm called AN-RWLLE ( Locally Linear Embedding Algorithm Based on Adaptive Neighborhood and Reconstruction Weight) is proposed. Firstly, the local neighborhoods of each sample point are identified by calculating the cosine similarity of high-dimensional sample points, followed by an adaptive selection of the optimal neighborhood within those neighborhoods. Secondly, the distance features and structural features of the sample points within the optimal neighborhood are combined to thoroughly explore the manifold structure of high- dimensional data and achieve weight reconstruction. Lastly, support vector machines are employed for feature recognition, preserving the intrinsic characteristics of high-dimensional data in a lower-dimensional space. Experimental results demonstrate that the AN-RWLLE algorithm exhibits excellent visualization, clustering performance, and effective feature extraction capabilities on two sets of bearing fault datasets.

Key words:  local linear embedding, feature extraction, adaptive neighborhood, reconstruction weight, bearing fault diagnosis