Research On Belief Function Based Decision Fusion For Wireless Sensor Networks

In recent years, Wireless Sensor Networks (WSNs) have been successfully applied in applications like industrial and environment monitoring, smart home and medical care. Driven by the constant improvement of sensor’s capacity in sensing, storage and computation, nowadays WSNs are becoming more and more intelligent, heterogeneous and massive-data. Decision fusion technique is a hot topic in WSN data processing due to its advantages in improving the data processing accuracy and efficiency. At the same time, it significantly reduces the data transmission amount, which greatly prolongs network’s operation lifetime.This paper gives a detailed introduction on WSN decision fusion techniques and belief function theory, along with some existing decision fusion approaches. Then the belief function theory based decision fusion algorithms in WSNs is studied, and all the proposed fusion algorithms are validated by experiments and comparisons. The main contributions include the following three aspects:(1) For the decentralized target classification problem in WSNs, a belief function theory based reliability-probability (RP) decision fusion algorithm is proposed. Existing basic belief assignment (BBA) constructing algorithms just unilaterally considers the classifiers’ epistemic uncertainty of or decisions’ aleatory uncertainty, this paper regards belief measurement as the lower bond of probability measurement, and presents an RP BBA construction algorithm under simple support belief function framework. By using a κ-nearest distance based reliability estimation algorithm, local BBAs are obtained via the proposed RP BBA construction algorithm in company with conditional probabilities of decisions. At last, the explicit BBA expression in fusion center is derived and a simple RP fusion rule is obtained. As a result, the complex Dempster combination operation is avoided, and the data transmission amount is decreased.(2) For the distributed target detection problem in WSNs, a corresponding RP fusion based target detection rule is proposed. By considering the difference between the probability density functions (pdfs) of target signal and unrelated noise, a cumulative probability density function (cdf) based reliability estimation algorithm is presented. BBAs in local sensors are constructed by combining with the decision reliability, probability of detection and probability of false alarm. At last, the RP fusion rule in binary target detection scenario is derived. Compared with existing belief function based fusion schemes, the proposed rule has advantages in easy implementation, lower data transmission amount, and it is compatible with commonly used threshold based target detection schemes.(3) Validate the proposed fusion algorithm by simulations and comparisons. In decentralized target classification, both random synthesize dataset and real vehicle surveillance data set are used to test the performance. The results prove that, no matter how local classification accuracy or sensor number changes, the proposed rule always performs better than the Naive Bayesian fusion and weighted majority voting rule. In distributed target detection, Monte Carlo simulation is used to test the performance and experimental results demonstrated that, compared with existing belief function based fusion, CV fusion and hard decision fusion schemes, the proposed algorithm always has much better performance in detection performance, diversity gains and anti-fading.