Diagnosis and prognosis processes are necessary to optimise the dependability of systems and ensure their safe operation. If there is a lack of information, faulty conditions cannot be identified and undesired events cannot be predicted. It is essential to predict such events and mitigate risks, but this is difficult in complex systems.Abnormal or unknown faults cause problems for maintenance decision makers. We therefore propose a methodology that fuses data-driven and model-based approaches. Real data acquired from a real system and synthetic data generated from a physical model can be used together to perform diagnosis and prognosis.As systems have time-varying conditions related to both the operating condi- tions and the healthy or faulty state of systems, the idea behind the proposed methodology is to generate synthetic data in the whole range of conditions in which a system can work. Thus, data related to the context in which the system is operating can be generated.We also take a first step towards implementing this methodology in the field of rolling element bearings. Synthetic data are generated using a physical model that reproduces the dynamics of these machine elements. Condition indicators such as root mean square, kurtosis and shape factor, among others, are calculated from the vibrational response of a bearing and merged with the real features obtained from the data collected from the functioning systemFinally, the merged indicators are used to train SVM classifiers (support vector machines), so that a classification according to the condition of the bearing is made independently of the applied loading conditions even though some of the scenarios have not yet occurred.