Virtual sensor approaches for condition monitoring of vertical transportation systems
- DIRECTORS: Wim Desmet, Jan Croes y Óscar Salgado
- UNIVERSITY: KU Leuven
Condition monitoring of electro-mechanical systems is receiving increasing attention, motivated by the growing demands on cost efficiency, reliability and quality. Condition monitoring strategies aim at identifying the system’s state of health, to detect and diagnose faulty conditions and early degradation. This information could then be used to predict optimal maintenance timing, reducing operational costs and increasing safety. To this end, advanced signal processing methods are used to extract fault-sensitive features from system measurements. However, it is often the case that the required sensors are not available, due to economic, operative or safety reasons. In such situations, virtual sensors are a valuable solution to estimate variables of interest that can not be measured.
A virtual sensor is a signal which is not measured, but rather inferred from the available measurements. An attractive solution to obtain virtual sensors is to continuously update a physics-based model using measurements and statistical inference. Then, the virtual sensors are derived from this updated model. This method takes advantage on the knowledge of the system’s physics, thus avoiding the requirement of costly measurement campaigns to describe the system’s behavior experimentally. However, developing and setting-up complex models is still a task for experts, and thus model and simulation is relegated to the system’s design phases. Moreover, models for monitoring purposes often have more strict limitations than design models, for instance, in terms of identifiability and computational cost. Thus, to fully exploit model- based strategies during the system’s operation, new synergies are required between modeling and monitoring. Models should be made adaptable to fit the monitoring requirements, and monitoring algorithms should be as generic as possible, and suitable for a wide range of models.
This thesis aims at facilitating the development of model-based virtual sensors, and their exploitation in a condition monitoring context. To this end, a framework based on acausal modeling tools and an industrial model exchange standard, the Functional Mock-up Interface (FMI), is discussed as the most suitable approach. On the one hand, acausal modeling tools promote model re- usability and scalability. On the other hand, the FMI facilitates the integration of models with monitoring algorithms, protecting the model intellectual property and making the framework independent from the modeling source. In particular, this thesis uses non-linear state estimation algorithms to update the model’s predictions. Namely, an augmented Extended Kalman Filter (EKF) is used to jointly estimate states and unknown inputs/parameters. In addition, in the context of state estimation, a method is proposed to select the model abstraction and the sensor set required by the estimator. This method builds upon the presented framework’s flexibility to create different models.
This is an industrially driven thesis, and thus the presented contributions are applied to the monitoring of a relevant electro-mechanical system, an elevator. With the proposed model-sensor selection method, it is shown how friction forces can be estimated using only readily available sensors. The methodology is validated experimentally using a scaled test bench of the elevator. In addition to the estimated states and unknown inputs, virtual sensors of other subsystems are derived, including the tension forces in the cable and the cabin’s acceleration. The estimation results are used to extract fault-sensitive features, demonstrating the potential of model-based virtual sensors to monitor the condition of complex systems.