Model based sensor fusion with fuzzy clustering

Redundancy in multisensor systems can often be exploited to increase sensor accuracy and reliability. This can be done by sensor fusion techniques. Two of the most important fusion methods are the Kalman filter and weighted averaging. Both methods use simple linear models which is not appropriate when the sensors are correlated in a nonlinear way. The Kalman filter moreover requires knowledge about the signal statistics which is usually unavailable. Our new fusion technique involves two steps: In the first step a fuzzy model of the functional dependence between the sensor signals is generated using fuzzy c-elliptotypes clustering. In the second step the noisy sensor signals are fused by a projection onto the model. When the model is linear this fuzzy model based technique is equivalent to weighted averaging. But by using several local linear models it can also deal with nonlinear correlated sensors. In the experiments fuzzy model based sensor fusion reduced the sensor noise error by 3% to 11%