Novel Graph-Based Features for Bearing Fault Diagnosis: Two Aspects of Time Series Structure

Abstract

The feature-based methods for bearing fault diagnosis in prognostics and health management have been achieved satisfactory performances because of their robustness to noise and reduced dimension by pre-defined features. However, widely employed time- and frequency-domain features are insufficient to recognize the global pattern that indicates the structure of a time-series instance. In this paper, we propose two novel graph-based features which reflect the connection strength and degree of time series, respectively. First, we construct a graph of which an edge is defined as the Euclidean distance between each pair of time steps to measure the strengths of connections between the nodes. The other graph is constructed by the visibility algorithm, which converts a time series into a complex network to reflect the degrees of connections. Then, we calculate the Frobenius norms of the adjacency matrices of both graphs and use them as features for bearing fault diagnosis. To verify the proposed features, we performed several experiments with both synthetic and real datasets. From the synthetic datasets, it is observed that the changes in amplitudes and frequencies are detected by the features for the connection strength and degree, respectively. In addition, the proposed features also well-separate the distributions of each bearing state, including normal and several fault types, and show significant performance improvement as applied to the fault diagnosis task.