Abstract
With the continuous development of wireless communication technology, the wireless electromagnetic environment is increasingly complex, which results in the difficulty of modulation recognition of communication signals. In this paper, combining the advantages of ResNet and DenseNet, we propose a blind modulation recognition model based on deep learning. In this model, we reduce the two-dimensional convolution neural network in ResNet into the one-dimensional convolution neural network and then embed it into DenseNet. The identity mapping of ResNet and the dense connection of DenseNet, which strengthen feature propagation and encourage feature reuse and reduce the number of parameters, make the model take full advantage of multi-layer features to improve the ability of feature extraction and reduce the computational complexity. The experimental results on the RadioML2016.10b dataset show that the recognition accuracy of the model can reach 93.7% at high SNR.
Introduction
In modern wireless communication, the electromagnetic environment is becoming increasingly crowded and complex. For applications such as cognitive radio, spectrum monitoring, and electronic warfare, Automatic Modulation Recognition (AMR) is a critical task. AMR involves identifying the modulation scheme of a received signal without prior knowledge of the transmission parameters.
Traditional AMR methods generally fall into two categories: likelihood-based and feature-based. Likelihood-based methods offer high accuracy but are computationally intensive and sensitive to channel estimation errors. Feature-based methods are more robust but rely heavily on manual feature engineering. Deep learning has emerged as a powerful alternative, yet many existing models utilize 2D Convolutional Neural Networks (CNNs) designed for images, which are often inefficient for the 1D temporal nature of radio signals.
System Model
The general framework for an AMR system involves a transmitter, a wireless channel subject to noise and fading, and a receiver that performs signal processing to identify the modulation type.
Proposed Hybrid Architecture
This research proposes a hybrid deep learning model that integrates the strengths of ResNet and DenseNet while tailoring the operations for signal processing.
1D Convolutional Shift
The primary innovation is the transition from 2D to 1D convolutions. Since radio signals are time-series sequences, 1D filters are more naturally suited to extract temporal features from raw I/Q (In-phase and Quadrature) data compared to standard image-based 2D filters.
Model Structure
The architecture consists of an input layer, six specialized ResNet blocks, and dense layers for classification.
- ResNet Influence: Within each block, "identity mapping" (shortcut connections) allows the network to bypass layers, mitigating the vanishing gradient problem and facilitating the training of deeper networks.
- DenseNet Influence: The blocks are interconnected using a dense scheme. While early connections utilize ResNet-style addition, subsequent stages employ DenseNet-style concatenation. This architecture encourages "feature reuse," ensuring that every layer can access features learned by all preceding layers.
- Optimization: The model incorporates Batch Normalization (BN) for training stability and Dropout to mitigate overfitting. The Adam optimizer is used for efficient weight updates.
Experimental Evaluation
The model was evaluated using the RadioML2016.10b dataset, which includes 10 modulation modes (8 digital and 2 analog) simulated under realistic channel conditions including multipath fading and additive white Gaussian noise (AWGN).
Performance Metrics
- Accuracy: The model achieved a peak recognition accuracy of 93.7% at an SNR of 16dB.
- Comparison: It outperformed several baseline models, including standard CNNs (87.4%) and CLDNNs (88.5%).
- Modulation Specifics: The model showed exceptional performance on high-order modulations like QAM16 (97.6%) and QAM64 (97.3%), which are typically challenging for simpler architectures.
- Efficiency: A key advantage of this hybrid approach is its computational efficiency. The model requires significantly fewer parameters and floating-point operations (FLOPs) compared to standalone ResNet or DenseNet implementations.
Discussion and Conclusion
While the model performs exceptionally well on digital modulations, it faces challenges with certain analog signals. Specifically, Wideband Frequency Modulation (WBFM) accuracy remains below 50%, often being confused with AM-DSB. This indicates that current feature extraction techniques may need further refinement to capture the specific characteristics of analog modulation.
Additionally, performance degrades at low SNR levels (below -6dB), a common limitation across deep learning-based AMR models. Future research will explore specialized layers for analog signals and the integration of attention mechanisms to further enhance robustness in high-noise environments.
Summary updated