Hybrid Neural Network Classifying Recognisable and Detecting Unrecognisable Radio Signals

A. Iversen, N.K. Taylor, K.E. Brown (UK), and J. Kårstad (Norway)


Classification, detection, neural network, radio communication


The Multi-layer Perceptron (MLP) classifier is widely used in classification tasks but has been shown to be inadequate in handling unrecognisable data, i.e. data that does not belong to any of the classes known to the classifier. In this paper, we look at modulation recognition of radio communication signals where the classifier is likely to be exposed to both recognisable signal types (which it has been trained on) and unrecognisable signal types (which it has not been trained to classify). We address this problem proposing a novel hybrid MLP classifier. Experimental results show that the hybrid classifier classifies recognisable signals very well whilst at the same time detects a large proportion of unrecognisable signals and thus prevents these from being wrongly classified. The hybrid classifier outperforms standard MLP classifiers incorporating output thresholding techniques for detecting unrecognisable data.

Important Links:

Go Back