Encoder-based weak fault detection for rotating machinery using improved Gaussian process regression

Abstract

Encoder signal analysis has proven to be a novel and cost-effective tool for the health monitoring of rotating machinery. Nevertheless, how to effectively detect the potential fault utilizing encoder information, especially at an early stage, remains a challenging issue. In light of this limitation, an improved Gaussian process regression analysis is proposed for the weak fault detection of rotating machinery via encoder signal. In this article, the Gaussian process regression model is first introduced to estimate the instantaneous angular speed and its confidence interval. Subsequently, to improve the robustness of Gaussian process regression under weak fault conditions, a spectral density complex kernel is constructed through modeling the spectral density with a mixture of Gaussians. Finally, built upon the eigenvalue decomposition, the optimal inference approach of improved Gaussian process regression is proposed. Compared with other regression methods, the major contribution is that the new method not only enhances the weak fault-related features but also sets their confidence interval adaptively. Using the proposed improved Gaussian process regression, the interference components are suppressed, while the fault-related instantaneous angular speed outliers are accurately detected. In addition, the significance of fault can be quantitatively evaluated according to the confidence level of the improved Gaussian process regression. The simulated and experimental analyses manifest that the proposed improved Gaussian process regression method can effectively identify the early weak fault. It may offer an effective tool for early fault detection of rotating machinery in industrial applications.