Abstract
In this work, we investigate the value of employing deep learning for the task of wireless signal modulation recognition. Recently, a framework has been introduced by generating a dataset using GNU radio that mimics the imperfections in a real wireless channel and uses 10 different modulation types. Further, a convolutional neural network (CNN) architecture was developed and shown to deliver performance that exceeds that of expert-based approaches. Here, we follow this framework and find deep neural network architectures that deliver higher accuracy than the state of the art. We tested the baseline architecture and found it to achieve an accuracy of approximately 75%. We first tune the CNN architecture and find a design with four convolutional layers and two dense layers that gives an accuracy of approximately 83.8% at high SNR. We then develop architectures based on Residual Networks (ResNet) and Densely Connected Networks (DenseNet) to achieve high SNR accuracies of approximately 83.5% and 86.6%, respectively. Finally, we introduce a Convolutional Long Short-term Deep Neural Network (CLDNN) to achieve a peak accuracy of approximately 88.5% at high SNR.
Introduction
In wireless communication, modulation is the process of varying a carrier signal to transmit information. For a receiver to function correctly, it must first identify which modulation scheme (e.g., QPSK, QAM16) is being utilized. Traditionally, this required "expert-based" approaches relying on prior knowledge of signal and channel parameters. However, as wireless environments become more complex and dynamic, there is a growing need for autonomous modulation recognition. While basic CNNs have shown promise, they often struggle with vanishing gradients as they get deeper and fail to fully capture the temporal (time-based) nature of radio signals.
Methods and Architectures
The authors evaluated and optimized four distinct deep learning architectures using the RadioML2016.10b dataset. This dataset consists of 160,000 samples of 10 different modulation types (8 digital, 2 analog) with Signal-to-Noise Ratio (SNR) levels ranging from -20dB to +18dB.
- Tuned CNN: An optimized version of earlier models, featuring four convolutional layers followed by two dense layers.
- Residual Networks (ResNet): These use "shortcut paths" to skip layers, allowing the network to learn residual mappings and mitigating the vanishing gradient problem.
- Densely Connected Networks (DenseNet): Each layer receives inputs from all preceding layers, maximizing feature reuse and improving information flow.
- Convolutional Long Short-term Deep Neural Network (CLDNN): This hybrid model combines CNN layers for spatial feature extraction with Long Short-Term Memory (LSTM) units designed to recognize long-term temporal correlations inherent in time-domain radio signals.
Key Insights and Results
The research demonstrated a clear hierarchy of performance at high SNR. While the baseline CNN achieved ~75% accuracy, the tuned CNN and ResNet reached ~83.8% and ~83.5% respectively. DenseNet further improved this to 86.6%, and the CLDNN emerged as the top performer with 88.5% accuracy.
The Depth Trap and Temporal Importance
A significant finding was the "depth trap" in standard CNN architectures. Simply adding more layers to a standard CNN eventually leads to performance degradation. As illustrated below, a 4-layer CNN was found to be the optimal configuration for this task; beyond this point, training becomes unstable without the use of residual or dense connections.
Furthermore, the superior performance of the CLDNN suggests that radio signals are not merely "images" of data. Their sequential, causal nature means that memory-based units like LSTMs are significantly better at extracting relevant features than pure spatial filters.
Limitations and Future Work
Despite these advancements, models still struggle to distinguish between highly similar constellations, such as QAM16 and QAM64, which are vulnerable to even small amounts of noise. Additionally, "silence periods" in analog voice signals (WBFM and AM-DSB) lead to high misclassification rates as the signals appear identical to the model during these intervals. Future work may involve specialized pre-processing, longer observation windows, or optimizing the trade-off between the high accuracy of models like CLDNN and the computational latency required for real-time edge devices.
Summary updated