System Identification of an Engine-load Setup Using Grey-box Model
With the demand for more comfortable cars and reduced emissions, there is an increasing focus on model-based system engineering. Therefore, developing accurate vehicle models has become significantly important. The powertrain system, which transfers the engine torque to the driving wheels, is one of the most important parts of a vehicle. Having a reliable methodology, for modeling and parameter estimation of a powertrain structure, helps predict different kinds of behaviors such as torsional vibration which is beneficial for a number of applications in automotive industry. Examples of such cases are ride quality evaluation and model-based fault detection.
This thesis uses the knowledge from the system identification field, which introduces the methods of building mathematical models for dynamical systems based on experimental data, to model the torsional vibration of an engine-load setup. It is a subsystem of the vehicular powertrain and the main source of vibration is the engine fluctuating torque. The challenges are handling a more complicated model structure with a greater number of unknown parameters as well as showing the importance of data information for acquiring better identification performance. Since the engine-load setup is modeled physically here, its state-space equations are available and a grey-box modeling approach can be applied in which the well-known prediction error method is used to estimate the unknown physical parameters. Moreover, a structural identifiability analysis is performed which shows that all of the model parameters are identifiable assuming informative input.
Two main aspects are considered to present an appropriate modeling methodology. The first is simplification of the model structure according to frequency range of interest. This is achieved by performing modal shape analysis to obtain how many degrees-of-freedom are necessary at different frequency ranges. The results show that a 7 degrees-of-freedom model can be simplified to a 2 degrees-of- freedom structure and still have the desired performance for a specific application such as misfire detection.
The second aspect concerns using an appropriate data set, which has the required information for estimation of the unknown parameters. By analyzing the simulation data from a known system, it is shown that the parameters of the 2 degrees-of-freedom model can not be estimated accurately using measurements from a normal combustion data set. However, all the parameters except damping coefficient converge to their true values by using a data set which has misfire in the input torque from the engine. A high estimation variance plus flat loss function indicate that the damping coefficient has no significant influence on the model output and consequently can not be estimated correctly using the available measurements. Thus, to increase the accuracy of the results during estimation on real data, the damping coefficient(s) is assumed to be known. Both the 2 and 7 degrees-of-freedom models are validated against a fresh data set and it is shown that the simulated output captures the important parts of the actual system behavior depending on the application of interest.
Senast uppdaterad: 2021-11-10