Automatic modulation classification is a challenging problem with multiple applications including cognitive radio and signals intelligence. Most of the existing efforts to solve this problem are only applicable when the signal to noise ratio (SNR) is high and/or long observations of the signal are available. Recent work has focused on applying shallow and deep machine learning (ML) to this problem. In this paper, we present an exploration of such deep learning and ensemble learning techniques that was used to win the Army Rapid Capability Office (RCO) 2018 Signal Classification Challenge. An expert feature extraction and shallow learning approach is discussed in a simultaneous publication. We evaluated multiple state-of-the-art deep learning network architectures and adapted them to work in the RF signal domain instead of the image/computer-vision domain. The best deep learning methods were merged with the best expert feature extraction and shallow learning methods using ensemble learning. Finally, the ensemble classifier was calibrated to obtain marginal gains. The methods discussed are capable of correctly classifying waveforms at-10 dB SNR with over 63% accuracy and signals at +10 dB SNR with over 95% accuracy from an Army RCO provided training set.

Automatic modulation classification is a challenging problem with multiple applications including cognitive radio and signals intelligence. Most of the existing efforts to solve this problem are only applicable when the signal to noise ratio (SNR) is high and/or long observations of the signal are available. Recent work has focused on applying shallow and deep machine learning (ML) to this problem. Feature generation, where raw signal information is transformed prior to attempting classification is a key part of this process. A big question that researchers face is whether to let the deep learning system infer the relevant features or build expert features based on expected signal characteristics. In this paper, we present novel signal feature extraction methods for use in signal classification via ML. The deep learning and combined approaches are discussed in a simultaneous publication. Expert features were utilized via ensemble leaning and shallow neural networks to win the Army Rapid Capability Office (RCO) 2018 Signal Classification Challenge. The features include both standard statistical measurements such as variance and kurtosis, as well as measurements tailored for specific waveform families. We discuss the best statistical descriptors along with a ranked list of signal features and discuss individual feature importance. We then demonstrate our implementation of these features and discuss effectiveness in estimating different modulation classes. The methods discussed when combined with deep learning are capable of correctly classifying waveforms at-10 dB SNR with over 63% accuracy and signals at +10 dB SNR with over 95% accuracy from an Army RCO provided training set.

No abstract