Radar intra-pulse modulation classification using convolutional neural networks

This dissertation presents a detailed investigation into the classification of radar intra-pulse modulation schemes. Recent years have seen increased waveform diversity in radar systems which, while making many aspects of pulse analysis more challenging, have presented new opportunities and features for the _eld of classification. This dissertation aims to address the increasing difficulty of pulse classification through the use of modern machine learning techniques - more specifically, by utilising convolutional neural networks. A wide range of modulation schemes was considered and simulated with realistic imperfections to create a dataset that was as representative of real-world scenarios as possible. Data representations of varying levels of abstraction were analysed in order to investigate the effects of data formatting on the performance of various classifiers. A classifier which made use of manual feature extraction was evaluated against a series of convolutional neural network classifiers in order to establish whether improvements in classification accuracy and throughput could be realised. This study also presents research into the viability of classifying data that has been degraded by real transmitter and channel effects using classifiers trained entirely on simulated data. The operation of the tested classifiers is analysed, and parallels are drawn between the feature extraction steps in convolutional neural networks and conventional signal features. The primary research questions in this study are whether machine learning approaches are able to improve on non-machine learning based classification techniques, and which data representations are best suited to convolutional neural network based classification. Classifiers were tested across 28 classes of modulation, with signal-to-noise ratios uniformly distributed between -5 dB and 20 dB. It was found that substantial performance and stability improvements could be achieved when convolutional neural networks were used over the tested non-machine learning based classification technique. The most promising classifier made use of time-frequency representations as an input, and was able to achieve a classification accuracy of 98%, while exhibiting extreme robustness against noise and pulse imperfections