Neural Network Approaches to Data Fusion Calculation and Correction in Wireless Sensor Network – A Case Study Comparison between BPN, RBFN and GRNN

Abstract

Abstract This paper explores the quality monitoring experiment by three existing neural network approaches to data fusion in wireless sensor module measurements. There are ten sensors deployed in a sensing area, the digital conversion and weight adjustment of the collected data need to be done. This method can improve the accuracy of the estimated data. In this study, the first neural network approach used to the correction of the sensed data is common Back Propagation Network (BPN). Although this approach leads to a quick convergence, the results are not the expectations and the training data output can't achieve a precise correction of the sensed data. The second approach is Radial Basis Function Network (RBFN), which can effectively achieve a correction of the sensed data and its training data output meets better the actual temperature measurements. The last approach is General Regression Neural Network (GRNN). GRNN and RBFN are both probabilistic neural networks. According to the experimental analysis of this study, GRNN can more accurately and quickly obtain the actual situation. The activity function of the nodes in hidden layers of GRNN adopts Gaussian function and thus GRNN has partial approximation ability. That's why GRNN can learn quickly. Besides, man-made adjustable parameter in GRNN only has one threshold so that the impact on the predicted results caused by man-made subjective assumptions can be avoided and the advantage of fast calculation can help network model be predicted more quickly.