A Fault Diagnosis Methodology Based on Non-Stationary Monitoring Signals by Extracting Features With Unknown Probability Distribution

Abstract

This paper studies features with the characteristic of unknown probability distribution, and its application on fault diagnosis based on non-stationary monitoring signals, which mainly consider the uncertainty as the main factor in masking fault diagnosis of practical industrial system. Generally, the probability distribution of the signal feature is unknown and prior information of the trend term is lacking. For this reason, different feature extraction methods, such as time-domain, frequency-domain and time-frequency-domain methods, have always been used to extract features, and they can be used to generate a high-dimensional and nonlinear initial feature set. However, the features' probability distribution is still unknown and prior information of the trend term is still lacking. In order to solve this top problem, Restricted Boltzmann Machine (RBM), with the advantage of feature learning and selection for initial feature set, has been stacked layer by layer to realize a high-dimensional nonlinear mapping between non-stationary signal features and fault modes. Two fault diagnosis experiments on self-confirmation sensor and rolling bearing shown the robustness and effectiveness of this proposed method.

Keywords

• Feature extraction
• restricted Boltzmann machine
• deep belief network
• non-stationary signals
• fault diagnosis