Toward Smart Condition Monitoring of Rotatory Machines: An Optimized Probabilistic Signal Reconstruction Methodology for Fault Prediction With Multisource Uncertainties

Faults in the critical components of a large rotatory machine often result in unplanned breakdowns, leading to a significant loss of property and life. Condition monitoring, as a key component in the smart maintenance of industrial equipment, has become a promising tool for providing automatic early alerting of potential damage to critical components, thus reducing potential outages and improving system safety and reliability while lowering maintenance costs. This is still a very challenging topic in various industrial fields because of data imperfection and multivariate correlation, as well as the variation in faults and components in different machines. This paper presents an optimized probabilistic signal reconstruction methodology to address these challenges in the fault prediction of rotatory machines using multivariate vibration signals. Three signal reconstruction methods, that is, Bayesian wavelet multiscale decomposition, probabilistic principal component analysis, and auto-associative kernel regression, were seamlessly integrated to address the noise, high dimensionality, and correlation in the sensed multivariate vibration data for accurate fault prediction. The bandwidth parameter in the auto-associative kernel regression approach was optimized to represent the health status of the rotatory machine. The obtained model was further utilized to predict the responses under unknown conditions. The alerting threshold, based on the squared mean errors of the predicted and measured time series, was automatically adjusted using a rolling window strategy and then employed to predict the possible fault. Finally, the validity, feasibility and generalization of the proposed methodology are illustrated by applying two cases of different rotating machines: centrifugal compressor and gas turbine.