Bayesian-Wavelet-Based Multisource Decision Fusion

Multisource information perception is the basic means for humans to explore the universe, and information decision fusion has become a crucial technique in some fields. Limited by the unknown distribution of multisource information, this article proposes a decision fusion method based on the distributed wavelet neural network (DWNN) and the Bayesian inference. The proposed fusion decision framework is a parallel network that consists of an empirical wavelet filtering layer, feature extraction layer, local decision layer, and decision fusion layer. Notice that the activation function of the feature extraction layer is a wavelet, and this nonlinear operation can be considered as a wavelet transform of the multisource data. Subsequently, an iterative learning method is adopted to minimize the estimated loss of the subnetwork and approximate the optimum decision model for local data. Furthermore, a decision fusion rule with the minimum Bayesian-like cost based on local evidence is adopted in the fusion center. Finally, a typical multisource decision fusion experiment of human surface electromyography (sEMG) and time series classification is presented to show the effectiveness of the proposed decision fusion structure.