EINDHOVEN UNIVERSITY OF TECHNOLOGY DEPARTMENT OF ELECTRICAL ENGINEERING Measurement and Control section PHILIPS CONSUMER ELECTRONICS SOUND REPRODUCTION TEAM

Electro-acoustical transducers with large dynamical range are needed for the reproduction of sound. For the reproduction of bass the transducers will have impractical large physical dimensions. Therefore it is important to find an alternative. One possibility is a small transducer which can be placed on an existing panel. The transducer consists of a mounted voice-coil and a free moving magnet. The main disadvantage of such a panel-transducer combination is its nonlinear behaviour. The main objective of this research project is to investigate the possibilities of using such systems. For compensating the nonlinear behaviour, system identification and modelbased control techniques will be used. Before a controller can be designed a mathematical model of the acoustical process has to be estimated. A literature study of the development and applications of such systems and the identification of nonlinear systems in general has been done. A prototype transducer was constructed. The dynamical behaviour of this nonlinear prototype has been analyzed and modelled using a black box approach. Therefor the NARMAX-modelset was chosen. The structure of the process has to be estimated to obtain a parsimonious model. For detecting the process structure and estimating the parameters simultaneously an orthogonal parameter estimation algorithm has been used. This algorithm has been implemented in MATLAB. The main goal has been defined as reducing the nonlinear behaviour of the acoustic transducer. For this state-feedback will be used. These states can be the measured process states or the states from a model. The acceleration of the transducer-membrane and the transducer input current can be considered as two of the process-states and will therefore be used as feedback signals. The process can be divided in three subprocesses. The input of these subprocesses is the transducer input voltage. The outputs are the current, membrane-acceleration and soundpressure respectively. Models of these several subprocesses have been estimated. Two main problems came into focus. The method is prediction error based which does not necessarily guarantee that the dynamics of the process will be modelled adequately. This has been shown by several simulations. This effect can be minimized by selecting a proper sample-frequency which will maximize the prediction-horizon. Due to a measurement problem, errors occurred in the dataset. Models have been fitted using a small dataset without these errors. Although these two problems obstructed the algorithm to fit a good model, it has been shown that the process identification using a combined structure-detection and parameter estimation method is very powerful.