Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Deep-Learned Compression for Radio-Frequency Signal Classification

Armani Rodriguez, Yagna Kaasaragadda, Silvija Kokalj-Filipovic

TL;DR

This work proposes a deep learned compression model, HQARF, based on learned vector quantization (VQ), to compress the complex-valued samples of RF signals comprised of 6 modulation classes, and assesses the effects of HQARF on the performance of an AI model trained to infer the modulation class of the RF signal.

Abstract

Next-generation cellular concepts rely on the processing of large quantities of radio-frequency (RF) samples. This includes Radio Access Networks (RAN) connecting the cellular front-end based on software defined radios (SDRs) and a framework for the AI processing of spectrum-related data. The RF data collected by the dense RAN radio units and spectrum sensors may need to be jointly processed for intelligent decision making. Moving large amounts of data to AI agents may result in significant bandwidth and latency costs. We propose a deep learned compression (DLC) model, HQARF, based on learned vector quantization (VQ), to compress the complex-valued samples of RF signals comprised of 6 modulation classes. We are assessing the effects of HQARF on the performance of an AI model trained to infer the modulation class of the RF signal. Compression of narrow-band RF samples for the training and off-the-site inference will allow for an efficient use of the bandwidth and storage for non-real-time analytics, and for a decreased delay in real-time applications. While exploring the effectiveness of the HQARF signal reconstructions in modulation classification tasks, we highlight the DLC optimization space and some open problems related to the training of the VQ embedded in HQARF.

Deep-Learned Compression for Radio-Frequency Signal Classification

TL;DR

This work proposes a deep learned compression model, HQARF, based on learned vector quantization (VQ), to compress the complex-valued samples of RF signals comprised of 6 modulation classes, and assesses the effects of HQARF on the performance of an AI model trained to infer the modulation class of the RF signal.

Abstract

Next-generation cellular concepts rely on the processing of large quantities of radio-frequency (RF) samples. This includes Radio Access Networks (RAN) connecting the cellular front-end based on software defined radios (SDRs) and a framework for the AI processing of spectrum-related data. The RF data collected by the dense RAN radio units and spectrum sensors may need to be jointly processed for intelligent decision making. Moving large amounts of data to AI agents may result in significant bandwidth and latency costs. We propose a deep learned compression (DLC) model, HQARF, based on learned vector quantization (VQ), to compress the complex-valued samples of RF signals comprised of 6 modulation classes. We are assessing the effects of HQARF on the performance of an AI model trained to infer the modulation class of the RF signal. Compression of narrow-band RF samples for the training and off-the-site inference will allow for an efficient use of the bandwidth and storage for non-real-time analytics, and for a decreased delay in real-time applications. While exploring the effectiveness of the HQARF signal reconstructions in modulation classification tasks, we highlight the DLC optimization space and some open problems related to the training of the VQ embedded in HQARF.
Paper Structure (8 sections, 5 equations, 6 figures)

This paper contains 8 sections, 5 equations, 6 figures.

Table of Contents

  1. Introduction
  2. System Model
  3. Hierarchical Vector-Quantized Compression of RF Datapoints
  4. HQARF Dataset for modulation recognition (ModRec)
  5. Generative Deep Learned Compression of RD datapoints using hierarchical VQ-VAE
  6. Neural Net architecture of VQ-VAE in HQARF
  7. Evaluation with the EfficientNet Classifier
  8. Conclusions and Future Work

Figures (6)

  • Figure 1: Shown is the information flow from a SDR through HQARF compression to layer 4 (L4). L4 requires the bandwidth $B^4,$ to store or transmit the compressed information about the datapoint $x$ composed of 1024 complex-valued samples, as opposed to directly storing/transmitting $x$ for a remote classifier to infer its modulation class. If a compressed representation $Z_{Q_i}, i \in \left\{0,\cdots,4\right\}$ is stored/ transmitted, the same HQARF model is used to recover $x,$ decompressing $Z_{Q_i}$ into $\hat{x}.$ We measure the effectiveness of compression by comparing the accuracy of R2FML between the reconstruction $\hat{x}$ and the original $x$ for various hierarchical compression rates;currently, $\xi= 1.37$.
  • Figure 2: The training of VQ-VAE - top: randomly initialized parameters of the encoder (E), decoder (D) and the $n_c$ quantization codebook (Q) vectors of dimension $\ell$; bottom: the final trained VQ-VAE where a single codeword's index from the trained Q will be associated with each of the $z_{e_n}=dim(z_e)[1]$ slices of $x$'s latent projection $z_e$. Therefore $x$ is compressed to $z_{e_n} \times \log_2(n_c)$ bits. On the right: The t-SNE visualization of the trained Q shows clusterization around a few codewords.
  • Figure 3: Table of the HAE encoders' input/ output dimensions and the SVD threshold (L5) for optimal $h$ parameters.
  • Figure 4: i/q scatterplot of 6 different classes based on the reconstructions across layers compared with the ideal (original) scatterplot. We concatenated 20 reconstructions of random datapoints of the same class, each comprised of 1024 complex-valued samples, and plotted them in the complex plane.
  • Figure 5: Spectrograms of L4 reconstructions and their originals for randomly selected modulations (0-5)
  • ...and 1 more figures