[Paper Review] Fully Dense Neural Network for the Automatic Modulation Recognition

Miao Du, Qin Yu+4 more|arXiv (Cornell University)|Dec 7, 2019

Wireless Signal Modulation Classification23 references4 citations

TL;DR

This paper proposes a Fully Dense Neural Network (FDNN) for automatic modulation recognition (AMR) that directly processes in-phase and quadrature (IQ) data without spectrogram conversion. By combining residual blocks, dense connections, and a Signal Attention mechanism, FDNN achieves higher recognition accuracy with significantly reduced model complexity compared to CNN and ResNet baselines on the RML2016.10a dataset.

ABSTRACT

Nowadays, we mainly use various convolution neural network (CNN) structures to extract features from radio data or spectrogram in AMR. Based on expert experience and spectrograms, they not only increase the difficulty of preprocessing, but also consume a lot of memory. In order to directly use in-phase and quadrature (IQ) data obtained by the receiver and enhance the efficiency of network extraction features to improve the recognition rate of modulation mode, this paper proposes a new network structure called Fully Dense Neural Network (FDNN). This network uses residual blocks to extract features, dense connect to reduce model size, and adds attentions mechanism to recalibrate. Experiments on RML2016.10a show that this network has a higher recognition rate and lower model complexity. And it shows that the FDNN model with dense connections can not only extract features effectively but also greatly reduce model parameters, which also provides a significant contribution for the application of deep learning to the intelligent radio system.

Motivation & Objective

• To address the limitations of traditional AMR methods that rely on expert-designed features or spectrogram preprocessing, which increase computational cost and memory usage.
• To improve recognition accuracy in automatic modulation recognition by directly processing raw IQ data using deep learning.
• To reduce model complexity and overfitting while enhancing feature extraction efficiency in AMR.
• To investigate the effectiveness of dense connectivity and attention mechanisms in end-to-end IQ-based modulation recognition.

Proposed method

• The FDNN architecture uses residual blocks to facilitate training of deep networks and improve gradient flow.
• Dense connections are implemented between all layers in residual blocks to promote feature reuse and reduce the number of trainable parameters.
• A Signal Attention (SA) mechanism is integrated into each residual block to recalibrate feature maps based on their importance, enhancing discriminative representation.
• The network is trained end-to-end on raw IQ data from the RML2016.10a dataset, avoiding intermediate spectrogram or constellation diagram representations.
• The architecture is designed to maintain high performance with minimal parameter count, improving generalization and reducing overfitting.
• Experiments compare FDNN against CNN, ResNet, DenseNet, and CLDNN models using standard metrics on the RML2016.10a benchmark.

Experimental results

Research questions

• RQ1 Can a fully dense neural network architecture achieve higher modulation recognition accuracy than conventional CNN and ResNet models when directly processing raw IQ data?
• RQ2 To what extent do dense connections reduce model complexity while maintaining or improving feature extraction capability in AMR?
• RQ3 How effective is the Signal Attention mechanism in recalibrating feature maps and enhancing recognition performance in IQ-based AMR?
• RQ4 Does the proposed FDNN architecture reduce overfitting compared to deeper networks like ResNet-50, despite its deeper structure?

Key findings

• FDNN achieved a higher average recognition rate than CNN2-IQ, ResNet-50, and ResNet-34 on the RML2016.10a dataset, with a 5.2% improvement over ResNet-50.
• The model reduced trainable parameters by approximately 80 times compared to the CNNR-IQFOC model, demonstrating significant efficiency gains.
• Despite having a deeper architecture, FDNN showed lower overfitting than DenseNet and CLDNN, as evidenced by better test set performance than validation set performance.
• The Signal Attention mechanism effectively recalibrated feature maps, with higher attention weights observed in early layers and shared importance across modulation types.
• The attention mechanism revealed that lower-layer features were more critical for discrimination, suggesting early layers capture more generalizable modulation characteristics.
• The combination of residual blocks, dense connections, and attention mechanisms led to superior performance with reduced model complexity and improved training efficiency.