Condition monitoring of rotating machinery elements, such as bearings, gears, ball screws, etc., is critical to ensure safe operation and optimal maintenance. Traditional condition monitoring solutions are primarily data-driven or machine learning (ML) based. These solutions face significant challenges due to their dependence on the availability of historical data and their limited transferability across different design and system configurations of rotating machinery. To address these limitations, this study proposes a novel multilayered digital twin framework for condition monitoring. The digital twin leverages physics-based models to generate synthetic data, overcoming the constraints of data availability. The proposed research uses lumped parameter models capable of simulating the dynamic response of the system for different fault severities and operating conditions. The dynamic response of both the virtual model and the physical system will be collected and processed to extract key features. Due to system variabilities, the virtual model will always have discrepancies with the response of the actual system. These discrepancies will be removed with the help of optimization models. The optimization model will determine the parameter tuning required, by minimizing the error between extracted features. Once tuned, the virtual model can be used to generate synthetic data for predefined fault conditions. Synthetic data can then be used to train ML-based models. These ML models utilize real-time responses from the machine element and facilitate real-time condition monitoring. To demonstrate the feasibility of this framework, initial validations using a ball screw feed drive system have been performed [1], [2]. The results show that by using the proposed digital twin concept, very accurate health conditions of the system can be predicted for different system configurations and other fleet variabilities even though historical data is not available.