Safety is more important than reliability or efficiency in railway, aerospace, oil & gas, and chemical industries. Regulations are very restrictive in sectors where safety is paramount. This makes maintainers replace critical components in initial stages of degradation, which implies a loss of useful life and a lack of information about advanced stages of degradation for those components. Nevertheless, this lack of data can be overcome using hybrid digital twins, also known as hybrid-model based approaches (HyMAs), which combine data-driven models with physics-based models. This fusion minimizes the occurrence of undesirable failures that may interrupt the functionality of critical systems in a safe or cost-efficient manner.

HyMAs have been studied at Luleå University of Technology by other Ph.D. students who found promising direction for future research in prognostics and health management (PHM) applications. Thus, this research work continues the direction defined in previous research with the proposal of HyMAs for a heating, ventilation, and air conditioning (HVAC) system installed in a passenger train carriage orientated to diagnostics and prognostics processes. The proposed hybrid modelling consists of the fusion of data obtained from two sources: data obtained from the real system and synthetic data generated by a developed physics-based model of the HVAC.

The HVAC system is considered a system of systems (SoS). Therefore, the physics-based model of the HVAC system is divided into four main systems: heating subsystem, cooling subsystem, ventilation subsystem, and cabin thermal networking subsystem. These subsystems are modelled considering the sensors installed in the real system and soft sensors, also known as virtual sensors, which provide crucial information for fault detection, diagnostics, and prognostics. These sensors defined in the physics-based model generate synthetic data which reproduce the behaviour of the system while a failure mode (FM) is simulated. Verification and validation are key processes to synchronise the response of the physics-based model with the signals obtained from the real system. Hence, the physics-based model is synchronised, verified, and validated using data collected by sensors located in the real system. These steps are conducted following guidelines suggested in the literature.

Different datasets containing real data and synthetic data while the HVAC system works in faulty and healthy states are used to train data-driven models for fault detection and diagnostics and to train data-driven models for prognostics.

Statistical features, such as shape factor, kurtosis, skewness, and sum square error, among others, are calculated from the selected signals. These features are labelled according to the related FMs and are merged with the features calculated from the data obtained from the real system. The data fusion is classified according to the condition indicators of the system in terms of FMs and level of degradation. The merged features are used to train data-driven models for fault detection and diagnostics. In addition, the real data can be loaded to the physics-based model to predict the degradation of the air filter.

Then, the prediction data are loaded to an exponential model that provides an estimation of the remaining useful life (RUL) of the air filter. To improve the prognostics model, the physics-based model is used to generate run-to-failure data which are used to train and test a deep convolutional neural network (CNN) which accurately estimates the RUL of the air filter.