Blind Source Separation Using Dictionary Learning in Wireless Sensor Network Scenario

The aim of this paper is to introduce a block-wise approach with adaptive dictionary learning for solving a determined blind source separation problem. A potential real-case scenario is illustrated in the context of a stationary wireless sensor network in which a set of sensor nodes transmits data to a multi-receiver node (sink). The method has been designed as a multi-step approach: The estimation of the mixing matrix, the separation of the sources by sparse coding and the source reconstruction. A sparse mixture from the original signals is used for estimating the mixing matrix, and later on, a sparse coding approach is used for separating the block-wise sources which are finally reconstructed by means of a dictionary. The proposed model is based on a block-wise approach which has the advantage of considerably improving the computational efficiency of the signal recovery process without particularly degrading the separation performance. Some experimental results are provided for comparing the computational and separation performances of the proposed system by varying the type of dictionary used, whether it is fixed or learned from the data.