PaperlayerView exact revision on arXiv

Source-linked AI summary

Real-Time Radio Technology and Modulation Classification via an LSTM Auto-Encoder

Ziqi Ke, Haris Vikalo

arXiv:2011.08295v1eess.SPcs.LG

TL;DR

Radio modulation and technology classification must remain accurate despite emitter diversity and channel effects, while also fitting low-cost computational constraints. The paper introduces an LSTM denoising auto-encoder that learns robust signal representations for classification. On realistic synthetic and over-the-air data, the framework generally outperforms competing methods while using substantially fewer computational resources.

  • Problem

    Growing emitter diversity and real-world channel effects make modulation and technology classification difficult, while existing methods may not meet low-cost platform accuracy and efficiency requirements.

  • Method

    An LSTM denoising auto-encoder corrupts and reconstructs received signals while learning low-dimensional representations used to infer modulation or technology type.

  • Results

    The framework generally achieves higher top-1 classification accuracy than competing methods while requiring significantly smaller computation resources.

  • Takeaways & Limitations

    The compact framework supports real-time classification of received signals on computationally constrained devices.

Abstract

from arXiv · show

Identification of the type of communication technology and/or modulation scheme based on detected radio signal are challenging problems encountered in a variety of applications including spectrum allocation and radio interference mitigation. They are rendered difficult due to a growing number of emitter types and varied effects of real-world channels upon the radio signal. Existing spectrum monitoring techniques are capable of acquiring massive amounts of radio and real-time spectrum data using compact sensors deployed in a variety of settings. However, state-of-the-art methods that use such data to classify emitter types and detect communication schemes struggle to achieve required levels of accuracy at a computational efficiency that would allow their implementation on low-cost computational platforms. In this paper, we present a learning framework based on an LSTM denoising auto-encoder designed to automatically extract stable and robust features from noisy radio signals, and infer modulation or technology type using the learned features. The algorithm utilizes a compact neural network architecture readily implemented on a low-cost computational platform while exceeding state-of-the-art accuracy. Results on realistic synthetic as well as over-the-air radio data demonstrate that the proposed framework reliably and efficiently classifies received radio signals, often demonstrating superior performance compared to state-of-the-art methods.

I. INTRODUCTION

Radio-signal classification supports spectrum allocation and interference mitigation, but real-world channel effects, growing emitter diversity, and resource constraints challenge accurate, efficient deployment. The paper proposes an LSTM denoising auto-encoder framework that learns robust representations for modulation and technology classification on compact devices.

  • Radio classification can inform spectrum allocation and radio-interference mitigation by identifying communication technologies and modulation schemes.
  • Expert feature-based methods may be difficult to generalize because crafted features may not capture all real-world channel effects.Likelihood-based methods have high computational complexity, while feature-based methods require carefully designed features and decision criteria.
  • Massive IQ-data acquisition creates storage and transmission burdens for distributed sensors, motivating more compact representations such as PSD features.
  • Existing deep-learning architectures are infeasible for distributed settings and low-cost computational platforms.
  • The proposed LSTM denoising auto-encoder corrupts received signals, reconstructs them, learns stable low-dimensional features, and classifies modulation or technology types.The framework uses amplitude and phase data for modulation classification and PSD data for technology classification.
  • The framework reports state-of-the-art modulation accuracy and significantly higher top-1 accuracy with a simpler structure, supporting real-time classification on compact devices.The paper demonstrates this capability using Raspberry Pi platforms.

A. Problem Formulation

The task is to identify a received signal’s modulation or technology type among K classes from sampled signal features. Modulation classification uses IQ data, while technology classification forms feature sequences from FFT-derived spectrum measurements.

  • A. Problem Formulation: Classification estimates P(y = C_k | r_ti) to identify a received signal’s modulation or technology type among K classes.C_k denotes the kth class and y is the signal’s true class.
  • A. Problem Formulation: Modulation classification represents each sampled signal with two-dimensional IQ features consisting of in-phase and quadrature components.RadioML2016.10A examples cover 11 modulation types at SNR = 10dB.
  • A. Problem Formulation: Pulse shaping, distortion, and other channel effects make modulation classes difficult to distinguish even for domain experts.
  • A. Problem Formulation: Technology classification scans candidate carrier frequencies, computes baseband FFTs, and concatenates average FFT coefficients into feature sequences.
  • A. Problem Formulation: Electrosense data span 50MHz to 6GHz with 0.1MHz scanning resolution and provide wireless magnitude spectra indexed by frequency and sampling time.
  • A. Problem Formulation: Evaluation considers top-1 accuracy over SNR, confusion matrices, trainable parameters, model size, and Raspberry Pi testing time.

B. An LSTM Denoising Auto-Encoder

The proposed classifier uses an LSTM denoising auto-encoder to learn compact representations from corrupted radio time series while simultaneously reconstructing signals and classifying modulation or technology type.

  • The LSTM processes radio signals as time series to capture temporal structure, producing hidden-state vectors from corrupted inputs.The encoder converts the corrupted signal into hidden states, while a shared fully connected layer serves as the decoder.
  • The denoising auto-encoder randomly masks input samples during training and uses the original signal for testing.The corrupted signal is fed to the auto-encoder, which learns from partially destroyed inputs.
  • The framework combines an LSTM denoising auto-encoder and classifier trained simultaneously for modulation and technology classification.The auto-encoder extracts representations while the classifier predicts the signal class.
  • The classifier uses the last hidden state, three fully connected layers, and a softmax output to predict the signal class.The LSTM auto-encoder has two layers, and the classifier is connected to the final hidden state.
  • The final objective is a weighted combination of mean-squared reconstruction loss and categorical cross-entropy classification loss.The balancing parameter λ lies in [0, 1]; the paper sets λ to 0.1 to promote reliable low-dimensional representations.
  • The reduced-dimensional hidden representation is intended to enable efficient classification with higher accuracy and significantly lower computational complexity.The model uses L2-normalized amplitude and normalized phase rather than IQ components.

C. Model Parameters

The model uses compact hidden states and regularized training settings, with optimization and architecture parameters selected for the denoising auto-encoder classifier.

  • Training uses Adam optimization, a minibatch size of 128, learning rate 0.001, and 150 epochs.These settings are reported for both tasks.
  • The LSTM hidden-state dimensionality is set to 32, whereas prior LSTM-based methods require more than 128 hidden states for their desired accuracy.Prior methods show significantly deteriorated accuracy with fewer hidden states.
  • The classifier fully connected layers contain 32, 16, and K nodes, while the decoder dense layer contains m nodes.K denotes the number of classes in the classifier architecture.
  • Dropout is set to 0.2 for the LSTMs and fully connected layers, and 10% of training-input entries are masked by zero.The masking supports the denoising-auto-encoder training procedure.

A. Performance Comparison on RadioML2016.10A

On RadioML2016.10A, the proposed denoising auto-encoder achieves the highest reported average top-1 accuracy while using fewer computational resources and classifying faster than competing models.

  • 61.72% average top-1 accuracy across all SNRs exceeds LSTM (60.49%), CLDNN (56.78%), and CNN (51.29%).
  • 1.1% higher accuracy across all SNRs results from training with noise rather than the original signal.
  • 91.55% average accuracy from 0dB to 18dB exceeds LSTM (90.26%), CLDNN (83.31%), and CNN (73.9%).
  • At low SNRs, confusion increases, including AM-DSB versus WBFM and QAM16 versus QAM64; the latter reflects QAM16 being a subset of QAM64.
  • The proposed model has the fewest trainable parameters, FLOPs, and memory requirements among the compared models.
  • On Raspberry Pi 4, the model classifies approximately 5.6×, 5.4×, and 1.9× faster than LSTM, CLDNN, and CNN, respectively.

B. Performance Comparison on RadioML2018.01A

On the realistic RadioML2018.01A dataset, the proposed model outperforms VGG and RN in classification accuracy while requiring fewer resources and achieving higher throughput.

  • 0.9% higher accuracy across all SNRs results from training with noise rather than the original signal.
  • At low SNRs, all considered models increasingly confuse GMSK, OQPSK, and BPSK signals.
  • The proposed model has the fewest trainable parameters, FLOPs, and memory requirements among the considered models.
  • On Raspberry Pi 4, it classifies approximately 2.4× faster than RN and 1.6× faster than VGG.

C. Performance Comparison on Electrosense Data

On over-the-air Electrosense PSD data, the proposed model slightly outperforms LSTM for technology classification and uses substantially fewer resources with higher Raspberry Pi throughput.

  • The proposed model performs slightly better than LSTM on Electrosense technology classification.
  • DVB and LTE are difficult to distinguish because their power spectra are highly similar and both use OFDM.
  • The proposed model has significantly fewer trainable parameters and requires fewer FLOPs and memory space than LSTM.
  • On Raspberry Pi 4, the model classifies approximately 4.3× faster than LSTM.
  • On Raspberry Pi 3, the model classifies approximately 6× faster than LSTM.

IV. CONCLUSIONS

The paper introduces an LSTM denoising auto-encoder that learns robust low-dimensional features for modulation and technology classification, combining reconstruction with inference.

  • The framework introduces a denoising auto-encoder for inferring modulation and technology types from received radio signals.
  • The LSTM auto-encoder learns stable, robust features from noise-corrupted signals and reconstructs the original signals while inferring their class.
  • Empirical studies generally show higher top-1 accuracy than competing methods while requiring significantly fewer computational resources.
  • The compact architecture can be implemented on affordable computational devices for real-time classification of received signals.
Loading 2011.08295v1…