Curved Exponential Family Manifold for Localization in Wireless Sensor Networks

Using statistical manifold theory in information geometry and natural gradient manifold learning localization method, the self-localization problem of wireless sensor networks based on received signal strength (RSS) was studied. First, a curved exponential family localization model was constructed according to probability density function. Then, aiming at the problem of locating unknown target nodes with given initial state values, combining gradient descent method, an optimal non-linear estimation method based on this model and its improved algorithm were proposed. The good properties of gradient descent method and simulation results show that these algorithms possess better convergence effect and higher positioning accuracy.