Health Monitoring of Rotating Machinery Using Probabilistic Time Series Model

Residual signal analysis is a promising tool for the health monitoring of rotating machinery (RM). The major challenge of this technique is how to establish an accurate time series model (TSM) such that the residual can highlight the incipient fault features. Traditional TSMs only provide a point estimation of the residuals, but fail to discriminate the fault impulses from noise and other disturbances. This deficiency prevents their wide applications in modern machinery, especially for those operating in harsh conditions. To tackle this issue, a new approach termed the probabilistic TSM (PTSM) is presented. In this approach, the recursive Gaussian process regression (GPR) is first introduced to explore the intrinsic dependency of time series from a probabilistic perspective. The merit of recursive GPR not only lies in its robustness to noise, but also it provides a confidence interval (CI) that indicates how likely the impulses are caused by the fault. Subsequently, for representing the health models from complex mechanical structures, an improved multiple kernel learning (MKL) method is constructed to ensure approximation accuracy. Finally, to improve the performance, the standard hyperparameters estimation method from the GPR framework is explored to determine the parameters. The results reveal a preferable capability to recover failure symptoms, even under strong noise interference.

Automated summary

Residual signal analysis is a promising tool for health monitoring of rotating machinery. The challenge lies in establishing an accurate time series model that can highlight fault features. The probabilistic TSM (PTSM) approach addresses this challenge by introducing recursive Gaussian process regression (GPR) and an improved multiple kernel learning (MKL) method. The results demonstrate the capability of recovering failure symptoms even under strong noise interference.