Digital Signal Processing Using Deep Neural Networks

Brian Shevitski,Yijing Watkins,Nicole Man,Michael Girard

DOI: https://doi.org/10.48550/arXiv.2109.10404

2021-09-22

Abstract:Currently there is great interest in the utility of deep neural networks (DNNs) for the physical layer of radio frequency (RF) communications. In this manuscript, we describe a custom DNN specially designed to solve problems in the RF domain. Our model leverages the mechanisms of feature extraction and attention through the combination of an autoencoder convolutional network with a transformer network, to accomplish several important communications network and digital signals processing (DSP) tasks. We also present a new open dataset and physical data augmentation model that enables training of DNNs that can perform automatic modulation classification, infer and correct transmission channel effects, and directly demodulate baseband RF signals.

Signal Processing,Machine Learning

What problem does this paper attempt to address?

The problem that this paper attempts to solve is: how to use deep neural networks (DNN) to perform various digital signal processing (DSP) tasks in the physical layer of radio - frequency (RF) communications. Specifically, the paper proposes a new DNN architecture specifically designed for the RF field, aiming to accomplish the following key tasks:

1. **Automatic Modulation Classification**: By training the model to recognize different types of modulated signals, accurate classification of the modulation types of unknown signals can be achieved.

2. **Transmission Channel Parameter Inference**: Directly infer the key parameters of the transmission channel from the received signals, such as phase offset, frequency offset, and signal - to - noise ratio, etc.

3. **Signal Denoising and Reconstruction**: Remove the noise in the received baseband signals through the auto - encoder network and reconstruct the originally transmitted signals.

4. **Message Attribute Classification**: For example, predict the number of symbols in the message.

5. **Direct Demodulation**: Directly demodulate digital information from the received baseband data without the need for the traditional multi - step demodulation process.

To achieve these goals, the authors developed a new open - data set and physical channel model to generate realistic RF signals for training DNNs. In addition, they proposed a hybrid model that combines auto - encoder convolutional networks and transformer networks, and this architecture can effectively handle multiple tasks in RF communications.

### Formula Summary

- **Transformation Formula for Channel Phase and Frequency Offset**:
\[
\begin{pmatrix}
I'(t) \\
Q'(t)
\end{pmatrix} =
\begin{pmatrix}
\cos(2\pi\Delta f t+\phi) & -\sin(2\pi\Delta f t+\phi) \\
\sin(2\pi\Delta f t+\phi) & \cos(2\pi\Delta f t+\phi)
\end{pmatrix}
\begin{pmatrix}
I(t) \\
Q(t)
\end{pmatrix}
\]

- **Rayleigh Fading Model**:
\[
x + iy=\frac{1}{\sqrt{N}}\sum_{n = 1}^{N}\left[\cos(2\pi\eta\tau\cos(\theta_n)+\alpha_n)+i\sin(2\pi\eta\tau\cos(\theta_n)+\beta_n)\right]
\]
where \(\eta = f_DT_M\), \(\tau=\frac{t}{T_M}\)

- **Total Loss Function**:
\[
\mathcal{L}=\sum_i\lambda_C^i\mathcal{L}_C^i(\theta_M)+\lambda_R\mathcal{L}_R(\theta_M)
\]
where \(\mathcal{L}_R(\theta_M) = MSE(Tx, Tx'(\theta_M))=|Tx - Tx'(\theta_M)|^2\)

- **Cross - Entropy Loss Function**:
\[
CE(L, L'(\theta_M))=-\sum_{j}v_j(L)\log(L'(\theta_M))
\]
where \(N_c\) is the total number of classes, and \(v_j(L)\) is the one - hot encoding vector of the true category.

Through these methods, the paper demonstrates the great potential of DNN in the RF communication field and provides a solid foundation for future research.