Application Of Decision Fusion For Sensor Data Quantization In Estimation Fusion And Decision Fusion With Fading Channels

In decentralized sensor networks, especially in wireless decentralized sensor net-works, since there exist various constrains such as power, channel bandwidth, thesensor data are usually compressed first, i.e., the sensor data are often quantizedinto a few bits before transmitted to the fusion center which makes final estimationand decision fusion with communication bandwidth limit. Therefore, how to findoptimal quantization becomes the attractive and di?cult problem in decentralizedsensor decision and estimation fusion. Aiming at this problem, this thesis analyzesthe relation between the decision fusion and estimation fusion on decentralized net-works at first, then proposes a new idea: converting an estimation fusion problemto decision fusion problem. Firstly, the interval in which the estimated parame-ter falls is partitioned into some subintervals, and a hypothesis is assigned to eachsubinterval. Then, from an estimation criterion, appropriate cost coe?cients forthe Bayes decision can be selected. Thus, we construct a mapping from estimationfusion to decision fusion so that the original estimation fusion is converted to a de-cision fusion. Hence, using the distribution knowledge of each sensor noise and themethod of decision fusion can get estimation fusion with quantized sensor data, i.e.,the outputs of decision fusion becomes the final estimation. Computer simulationshows that the performance of estimation fusion under such sensor data quantiza-tion is acceptable. Furthermore, this thesis made more discussion about suitablemethod to partition the interval, select cost coe?cients and fusion rule, etc. sothat the system can achieve better performance. Aiming at decision fusion problemwith fading channel in wireless sensor networks which has just been studied since recent a few years, this thesis studies the original decision fusion methods includingnecessary condition for optimal sensor rules and the iterative algorithm withouttransmission error, analyzes the relation between the transmission errors and fusionrule, extends the original methods to the channel fading case with the knowledgeof probabilities of channel transmission errors. Given a fusion rule for a networkdecision fusion problem, a necessary condition for optimal quantization of sensors'data is obtained, and an e?cient iterative algorithm is proposed without increasingcomputational complexity. Our results are di?erent from the results in existing lit-erature. Our methods don't require conditional independence between sensor data,nor do the independence of channels connecting sensors and fusion center. Thisthesis also solves the estimation fusion problem with fading channels by the resultsjust introduced above. At the end of this thesis, some computer numerical examplesare provided to support the theoretical analysis.