Multidimensional Dirichlet Process-Based Non-Parametric Signal Classification for Autonomous Self-Learning Cognitive Radios

In this paper, we propose a Bayesian non-parametric signal classification approach for spectrum sensing in cognitive radios (CR's). The proposed classification approach is based on the Dirichlet process mixture model (DPMM) that allows inferring the number and types of signals from their spectral and cyclic properties. The proposed algorithm is completely autonomous and does not require any prior knowledge of the existing signals or the number of distinct signal classes. We assume that the cluster parameters are drawn from a mixture model, where each mixture component parameterizes a specific observation model, including both Gaussian and non-Gaussian models. By using the Gibbs sampling, we estimate the observation model and cluster parameters that best fit the observed data. Given N data points, under certain regularity conditions, we derive an upper bound for the mean-squared error (MSE) in estimating the clusters means. A Bayesian prediction method is also developed to estimate the probability distribution of the data points. The proposed algorithm is applied to detect and classify WiFi and Bluetooth signals in the ISM band. Simulation results validate the proposed classification approach and show its robustness against channel impairments such as Rayleigh channel fading.