Learning Representations With Local and Global Geometries Preserved for Machine Fault Diagnosis

Recently, deep learning-based representation learning methods have attracted increasing attention in machine fault diagnosis. However, few existing methods consider the geometry of data samples. In this paper, we propose a novel method to obtain representations that preserve the geometry of input data. More specifically, we formulate two cost functions to preserve the local and global geometries of input data, respectively and another cost function to reconstruct the input data. Furthermore, to simplify the training process, we formulate a discrimination cost function based on the label information. By jointly optimizing all cost functions, the method can efficiently learn discriminative representations with the local and global geometry of input data preserved. Furthermore, the proposed method can obtain hierarchical representations without any additional tuning step. On two benchmark datasets, the proposed method demonstrates better fault classification performance and shorter training and test time. Therefore, it is an efficient tool to provide accurate information about machine conditions for making maintenance decision and saving costs.