IEEE Transactions on Cognitive Communications and Networking · 2023 · 31 citations · 29 references
ComputingComputer ScienceElectrical EngineeringFeature EngineeringMachine LearningTime Series AnalysisDeep LearningPattern RecognitionRadar Signal ProcessingRepresentation LearningTime Series ClassificationTime-frequency AnalysisConvolutional Neural NetworksDeep Learning-based ClassificationModulation ClassificationModulation RecognitionSignal-to-noise RatioSignalsTime-frequency Feature ExtractionTime-frequency Features
Existing work on modulation classification of radar signals widely employs two-stage algorithms of time-frequency transform (TFT) and convolutional neural network (CNN). This paper explores the vulnerability of the time-frequency representation (TFR) to noise in the traditional TFT-CNN framework. It proposes a deep TFT classification network (DTFTCNet) that does not require explicit TFT during testing. The network reformulates the cooperation scheme between TFT and CNN, proposes a deep TFT built into the neural network to replace conventional TFT, and obtains classification results and ideal TFR without TFT during testing, realizing an end-to-end deep learning framework. Experiments show that DTFTCNet outperforms state-of-the-art modulated recognition networks and conventional TFT-CNNs in nine radar signal datasets at all signal-to-noise ratio (SNR) nodes (−14 dB to 8 dB), and DTFTCNet can mimic various traditional TFT styles. The clarity of the TFR output from DTFTCNet is significantly better than the comparison method. This paper also aims to inspire researchers to explore cooperation between conventional TFTs and deep learning.
29