Communication signal generation and automatic classification with detection of unknown formats using neural networks

Abstract

The performance of algorithms for signal classification is often characterized according to their ability to correctly classify various signals belonging to a set of known signal formats. However, in a non-co-operative setting it may also be expected that the classifier will be presented with signals of unknown formats, i.e. signal formats not belonging to any of the categories in the training set. Typically, a neural network classifier will classify an unknown signal to the known signal format that it resembles the most. In EW applications, however, the ability to detect unknown signal formats is imperative. In this report a hybrid classifier scheme is proposed. The performance of this hybrid classifier is tested both with respect to its ability to correctly classify known signal formats as well as its ability to detect outliers. The results indicate that this two-stage classifier is able to detect most of the unknown signal formats, but at the expense of misclassifying some of the known signals. The trade-off between correct classification and detection of new signal formats can be adjusted by setting an appropriate CIV (Class Inherence Verification) threshold. In addition to discussing the classification algorithms and their performance the procedure for generation of test signals is also described.