SVM Classifier Approach to Enumerate Directional Signals Impinging on an Array of Sensors

The support vector machine classification method is used for source enumeration; i.e., estimating the number of sources contributed in generation of the signals received by the sensors of a passive array. The main motivation comes from the problems where the data model is not completely known, or the model is subject to some changes due to real environmental conditions, (i.e., the case of data model mismatch). With no model mismatch, because we know a-priori about the generative model of the data, the traditional statistical signal processing techniques yield better results compared to general machine learning techniques like SVM (which try to learn everything from training data samples, without any use of their generative data model, and also by employing a simple geometric structure). Monte-Carlo test results show the potential of the SVM to compete with statistical signal processing in this particular application. Specifically, the SVM has a slightly better performance in the case of model mismatch, (sensor gain perturbations, sensor phase perturbation, and for spatially correlated array noise), compared to traditional enumeration methods MDL and AIC, at very low SNR values