Unbalance faults are among the common causes of interruptions and unexpected failures in rotary systems. Therefore, monitoring unbalance faults is essential for predictive maintenance. While conventional time-invariant mathematical models can assess the impact of these faults, they often rely on proper assumptions of system factors like bearing stiffness and damping characteristics. In reality, continuous high-speed operation and environmental factors like load variations cause these parameters to change. This work presents a novel architecture for unbalance fault monitoring and prognosis, in which the bearing parameters are treated as variables that change with operating conditions. This enables the development of a more reliable mathematical model for continuous monitoring and prognosis of unbalance faults in rotor systems. This Bayesian inference framework uses Markov Chain Monte Carlo (MCMC) sampling to identify dynamic bearing parameters. Specifically, the Metropolis algorithm is employed to systematically evaluate the range of acceptable parameter values within the framework. A novel dual-MCMC loops explore and assess the parameter space, resulting in more accurate and reliable bearing parameter estimations. These updated parameters improve the demonstrated turbine rotor–bearing system’s unbalance assessment up to 74.48% of the residual error compared to models with fixed parameters. This validates the Bayesian framework for predictive monitoring and maintenance-oriented solutions.
In practical engineering applications, rotor–bearing systems inevitably exhibit inherent or residual faults such as imbalance and shaft-bow, originating from manufacturing tolerances, thermal deformation, or operational loading. Accurate monitoring of these faults and their evolution is fundamental to the effectiveness of modern prognostics and health management (PHM) frameworks. However, if such inherent faults are not identified at an early stage, substantial deviations in fault diagnosis may occur, thereby compromising the accuracy of subsequent prognostic assessments and maintenance strategies. This study presents a hybrid diagnostic methodology that integrates a physics-based model with neural network techniques to enhance rotor fault diagnosis. A Jeffcott rotor subjected to simultaneous disk imbalance and shaft-bow is used to demonstrate the methodology, and the results proves its superior capability for simultaneous fault identification. Nonetheless, discrepancies between model predictions and experimental results are observed, attributed to the presence of inherent faults within the rotor system. To address this issue, algorithms for inherent fault identification and compensation, supported by experimental verification, are developed. Following compensation, the accuracy in simultaneously diagnosing and estimating the parameters of imbalance and shaft-bow is significantly improved. The proposed methodology is designed for seamless integration into real-time monitoring systems of industrial rotating machinery.