On The Data Level, Feature Level Data Fusion Method

The theory and applications of data fusion are studied in this paper. The definition, fundamental, fusion levels, process model, fusion methods of data fusion are summarized firstly. According to the fusion levels, there are data level, feature level, decision level included in data fusion. Different levels has different characteristics, theory and algorithms. The theory. algorithms and applications of data level fusion and feature level fusion are mainly studied in this paper to explore effective fusion models ,relevant algorithms and applications in different fusion levels.In data level fusion, in allusion to observed noise in congeneric multi-sensor measure, the adaptive weighted fusion estimated algorithm of multi-sensor data is presented in this paper. Without of any pre-defined knowledge concerning sensors, the algorithm can adjust the fused sensor's weight according to changes in the sensor's variance, and the estimation error is kept as of least mean square. It is theoretically proved that this algorithm is linear and unbiased, and has least mean square errors. Simulation results show that the fused data are effective and much more sensitive ,accurate, reliable than the mean estimated algorithm.In feature level fusion, all kinds of information is being synthesized in fault diagnosis so as to obtain the integration of the running and faulty state, a method of which dada fusion is used in fault diagnosis is provided . In allusion to the function of associated memory ,classification optimized decision and rapid responding speed in practical applications, ANN is used in feature fusion. The fault diagnosis model of ANN in feature fusion is provided and applied in fault diagnosis example of electronic circuit. BP network, ameliorated BP network and RBF network are applied in simulation of fault diagnosis. The simulation results validate the strong fault diagnosis ability of ANN in feature fusion .The ability of three networks in fault diagnosis is analyzed and compared. The results show that RBF network is more suitable for fault diagnosis in feature fusion because of more rapid responding speed.