A Feature Extraction Method Based on Sparse Filtering with Local Structure Preserved and Its Applications to Bearing Fault Diagnosis

Abstract

Unsupervised feature learning, as a promising tool for extracting features automatically, overcomes shortcomings of traditional feature extraction methods which generally take plenty of effort on designing features. Among various unsupervised feature learning methods, sparse filtering is an efficient and popular one owing to its simplicity with fewer hyper-parameters. However, the standard sparse filtering fails to consider the local structure of input samples, which may affect and restrict its feature learning ability. To address this deficiency, a new feature extraction approach named sparse filtering with local structure preserved (SF-Local) is proposed. The objective function of SF-Local is composed by the standard sparse filtering and a local structural regularization (LSR) which is formulated to preserve the local structure of the input samples. To design LSR, we adopt complexity-invariant distance to select neighbors for each sample rather than the widely used Euclidean distance, which is beneficial to establishing a reliable K-nearest graph. Based on SF-Local, an intelligent fault diagnosis method is developed by leveraging a supervised classifier. The developed diagnostic approach is applied to diagnose the bearing faults as well as its performance is evaluated on a bearing data set with a constant operating condition and a data set with variable operating conditions. Experimental results demonstrate the effectiveness and superiority of the proposed SF-Local.