Ensemble Automatic Modulation Classification (EAMC)

In this proposal, we introduce EAMC, an Ensemble Automatic Modulation Classification approach. Our framework fuses three complementary heterogenous signal representations to alleviate the inherent challenges with radio frequency (RF) signal classification. In wireless communication systems, Automatic Modulation Classification (AMC) can be treated as a classification problem in Machine Learning (ML). Unlike images, RF signals change over time. Therefore, they have temporal characteristics in addition to spatial features, which cannot be ignored.ááConvolutional neural networks (CNNs) are superior in extracting spatial features by convoluting signals to generate useful high-level information. However, CNN models for the task of AMC are error prone under some uncertainty conditions such as different sampling rates. This vulnerability stems from the fact that CNNs only consider the spatial features and neglect the temporal attributes. Similarly, the Long Short-Term Memory networks (LSTM) are beneficial for exploring temporal correlations of signals and can learn gate weights to address the specified task by considering the previous cell state and input data. LSTM solves the problem of gradient explosion and the inability to learn long-term dependencies. However, an LSTM model for AMC may ignore the spatial features. Our solution will combine the CNN and LSTM for the task of AMC to enable the joint utilization of both temporal and spatial features.áEAMC adapts a CNN-LSTM three-stream structure, which considers the feature interaction and combines the advantages of CNN and LSTM. A parallel heterogeneous meta-learning method is proposed to combine In-phase and Quadrature (IQ), Amplitude and Phase (AP), and Constellation classifiers through a stacking approach into an ensemble classifier of the proposed heterogeneous base classifiers. By employing IQ, AP, and Constellation data at the same time, we reveal more discriminative features and allow the deep network to increase discrimination among different modulation patterns to further boost classification accuracy. Due to their heterogeneity, the three base classifiers are complementary in their per-class accuracy.á