The medical science branch and the development of medical ventilators is a fast growing area and the development of new techniques and products are important. Reducing the product development cycle time is therefore an important fact, since this could reduce the cost and the amount of personnel required to develop new products. Creating computer based test bench software of the medical ventilators instead of running tests on real machines is something that health care companies such as Getinge Maquet Critical Care are developing.

Medical ventilators and the corresponding treatments use compressed air and oxygen in order to assist or replace the spontaneous breathing procedure of a sick patient. This project focuses on the development of computer based models of the medical ventilator Maquet Servo-U and the desires of this project is to accurately simulate the high frequency pneumatic dynamics of the ventilator pipes and the attached components during the high frequency oscillation ventilation treatment (HFOV). The models that are investigated in this thesis focuses on an implementation based on system identification theory. By following this theory, mathematical models of dynamical systems are derived using the prediction error method (PEM). Two types of system identification modelling are investigated in this thesis, where the first type is called black box modelling which creates models that are based on inputoutput measurements. Auto regressive models with exogenous inputs (ARX), non linear auto regressive models with exogenous inputs (NLARX), auto regressive moving average models with exogenous inputs (ARMAX) and state space models are the black box models that are investigated in this thesis. The second type of modelling is called grey box modelling, that is models that requires some physical or initial knowledge about the system. In this project the model structure is based on the bond graph theory and by using the input-output measurements the grey box models are estimated. Both a linear structure and a non linear structure of the bond graph models are investigated.

The requirements of the estimated models are that the models should be able to give an estimate of the Y-piece pressure amplitude within an ±5%-error margin and the pressure bias within an ±2mbar-error margin, at a breathing frequency between 5-20 Hz. The estimated models are validated against some chirp signals with different amplitudes and bias flows. The estimated grey box parameters should also be reasonable close to the theoretical calculated parameters of the pipe system in order to be meaningful to use.

Two different setups of the ventilator system are considered. The first is a simplified setup that contains only the elastic pipes of the ventilator. The pipes are in this setup connected with one end to the inspiratory valve and the other end open. The second setup contains all pipes with the filter and humidifier attached. The pipes are in this setup connected at the inspiratory valve and at the expiratory valve.

The results shows that the estimated black box models generates better results than grey box models independent on the identification data. Some of the estimated black box models pass the requirements when the bias and amplitude of the validation data is quite similar to the bias and amplitude of the identification data and the frequency is around 15-20 Hz. However, when the amplitude and bias of the validation signal differs more from the identification signal or the frequency of the flow are outside the 15-20 Hz range the model fails. This leads to the conclusion that it is possible to estimate an accurate model for data that is close to the identification data in amplitude, bias and some frequencies, but that the estimated models does not cover all of the dynamics.

In order to aid these problems, the further work of this thesis mainly contains four things. The first one is to use an estimation data set that includes more of the dynamics of the pipe system, for example include more energy at low frequency, include different bias and amplitudes of the flow. The second one is to gain schedule the models depending on the bias and amplitude. This could give a higher ability to track different bias and amplitudes of the pressure curve. The third one is to change the cost function and make it react on the relative error, instead of the absolute error, in order to get a higher ability to track the low pressure amplitudes. The last one is an implementation based on Extended Kalman Filter (EKF) or Unscented Kalman Filter (UKF).