PaperlayerView exact revision on arXiv

Source-linked AI summary

Distributed Deep Learning Models for Wireless Signal Classification with Low-Cost Spectrum Sensors

Sreeraj Rajendran, Wannes Meert, Domenico Giustiniano, Vincent Lenders, Sofie Pollin

arXiv:1707.08908v2cs.NI

TL;DR

The paper asks whether wireless technology classification can operate at scale on distributed, low-cost sensors with limited bandwidth and computation. It proposes LSTM-based classifiers using time-domain amplitude and phase, and evaluates FFT-based and quantized models for resource-constrained deployment. The models achieve strong classification results, including close to 90% accuracy at high SNRs and 80% accuracy with averaged magnitude FFT data, while the authors identify remaining generalization and real-world variability limitations.

  • Problem

    Distributed low-cost sensors face limited computation, bandwidth, and costly IQ data transfer, motivating scalable wireless technology classification.

  • Method

    The paper develops LSTM classifiers from time-domain amplitude and phase, and studies averaged magnitude FFT inputs and quantized models for sensor deployment.

  • Results

    The proposed models achieve close to 90% accuracy at high SNRs, 80% accuracy using averaged magnitude FFT data, and 19% accuracy on the same 11-modulation dataset with indistinguishable magnitude spectra.

  • Takeaways & Limitations

    LSTM models can classify variable-length signals, while averaged magnitude FFT data supports basic technology classification under distributed sensor bandwidth constraints.

  • Takeaways & Limitations

    The models require further validation on varying real-world symbol rates and bandwidths, unknown channel conditions, and modulation parameters.

Abstract

from arXiv · show

This paper looks into the technology classification problem for a distributed wireless spectrum sensing network. First, a new data-driven model for Automatic Modulation Classification (AMC) based on long short term memory (LSTM) is proposed. The model learns from the time domain amplitude and phase information of the modulation schemes present in the training data without requiring expert features like higher order cyclic moments. Analyses show that the proposed model yields an average classification accuracy of close to 90% at varying SNR conditions ranging from 0dB to 20dB. Further, we explore the utility of this LSTM model for a variable symbol rate scenario. We show that a LSTM based model can learn good representations of variable length time domain sequences, which is useful in classifying modulation signals with different symbol rates. The achieved accuracy of 75% on an input sample length of 64 for which it was not trained, substantiates the representation power of the model. To reduce the data communication overhead from distributed sensors, the feasibility of classification using averaged magnitude spectrum data, or online classification on the low cost sensors is studied. Furthermore, quantized realizations of the proposed models are analyzed for deployment on sensors with low processing power.

I. INTRODUCTION

The paper addresses scalable wireless technology classification for distributed, low-cost sensors with limited computational and communication resources. It proposes LSTM-based AMC and evaluates FFT-based and quantized alternatives for sensor deployment.

  • Wireless technology classification supports spectrum enforcement, interference detection, and wireless environment analysis across distributed sensing networks.
  • The proposed LSTM classifier learns long-term temporal representations and handles variable input lengths without explicit expert feature extraction.
  • Limited sensor computation, costly IQ transfer and storage, and restricted bandwidth make large-scale AMC deployment difficult.
  • The paper studies classification from averaged magnitude FFT data, reducing communication cost by a factor 1000, and examines quantized models for low-cost sensors.
  • The contributions include time-domain amplitude-phase AMC, distributed classification using averaged magnitude FFT data, and quantized neural networks for sensor deployment.

II. PROBLEM STATEMENT

The problem frames modulation recognition as N-class inference from a received wireless signal affected by channel response and noise. The received signal is represented through in-phase and quadrature components derived from instantaneous amplitude and phase.

  • Technology or modulation recognition is formulated as an N-class classification problem using the received signal as the reference.
  • The received signal combines the transmitted signal, a time-varying channel impulse response, and additive white Gaussian noise.
  • The classifier estimates P(s(t) ∈ N_i|r(t)) for each class using the received signal r(t).
  • IQ representation expresses the received signal with I = A cos(φ) and Q = A sin(φ), where A and φ are instantaneous amplitude and phase.
  • The evaluated datasets assume sampling sufficiently high to capture the full signal bandwidth and include receiver imperfections.

III. MODULATION DATASETS

The study uses RadioML2016.10a as a baseline and extends it to test sample-rate dependence. The proposed model is introduced for time-domain amplitude and phase signals, alongside a CNN comparison.

  • The evaluation uses the publicly available RadioML2016.10a dataset as the baseline for training and testing.
  • RadioML2016.10a is synthetically generated with commercial modulation parameters and channel imperfections including frequency offset, sample-rate offset, noise, and multipath fading.
  • The baseline dataset contains signals with 4 samples per symbol and sample length 128.
  • The extended dataset varies samples per symbol and sample length, using 4 and 8 samples per symbol to evaluate robustness to changing symbol rates.
  • The paper introduces an LSTM model using time-domain amplitude and phase signals and details a CNN model for comparison.

A. LSTM primer

LSTM is presented as a recurrent model for time-series classification that learns persistent features through long-term dependencies and gating. In this paper’s classification setup, averaged magnitude FFT bins form a one-dimensional sequential input.

  • A. LSTM primer: LSTM is a recurrent neural network designed to learn long-term dependencies in time-series data.Its use is motivated by persistent feature learning from sequential signals.
  • A. LSTM primer: LSTM cells maintain an internal state and use input and forget gates to retain information over longer durations.The gating mechanism learns gate weights from previous state and input data.
  • A. LSTM primer: The averaged magnitude FFT representation uses one value per frequency bin as the sequential input to the model.Each input has dimension [n frequency bins, 1], with a_f denoting the averaged magnitude FFT bin at time t.
  • A. LSTM primer: The two-layer LSTM model supports modulation classification from either amplitude-phase signals or averaged magnitude-FFT signals.Figure 2 depicts the shared classification architecture for these two input representations.

B. Model for complex signals

For complex signal classification, the model processes normalized amplitude and phase sequences with stacked LSTM layers and maps the learned representation to 11 modulation classes. Its configuration and evaluation include varying layer depth, comparison with a CNN baseline, and practical TensorFlow deployment.

  • B. Model for complex signals: The amplitude-phase LSTM feeds normalized amplitude and phase values into two 128-cell LSTM layers before classifying 11 modulation schemes.The second layer’s 128-dimensional final output is passed to a dense softmax layer.
  • B. Model for complex signals: Amplitude is L2 normalized, while phase in radians is normalized between -1 and 1 before entering the LSTM.Both quantities are supplied as a two-dimensional vector at each time step.
  • B. Model for complex signals: The model’s classification behavior is analyzed across layer depths from 1 to 3 and cell counts from 16 to 256.These experiments examine how architecture choices affect performance.
  • B. Model for complex signals: A two-layer CNN with 8-tap filters serves as the baseline for subsequent comparisons.The baseline uses 256 filters in its first convolutional layer and 80 in its second.
  • B. Model for complex signals: Training and testing data are separated using mutually exclusive randomly generated indices.A softmax cross-entropy objective is used for discrete classification.
  • B. Model for complex signals: The trained TensorFlow model can be imported into GNU Radio for use with supported hardware frontends.Python and C++ bindings support portability to host-based SDR frameworks.

V. RESULTS AND DISCUSSION

The two-layer amplitude-phase LSTM achieves strong high-SNR modulation classification and outperforms standard machine-learning baselines on raw sensed data, while performance varies with SNR and modulation ambiguity.

  • Classification accuracy: 90% average accuracy was achieved by the two-layer LSTM across 0dB–20dB SNR.Layer depth saturated at two, which was selected for the final model.
  • Classification accuracy: Deep learning models outperformed SVM, random forest, k-nearest neighbors, and Gaussian Naive Bayes on the same amplitude-phase data.Random forest with 150 trees reached close to 70% at very high SNR, while other standard models reached only around 26%.
  • Confusion matrices: At 18dB SNR, the confusion matrix was sharply diagonal, with difficulty separating AM-DSB and WBFM.The paper attributes this difficulty mainly to silence periods in the real audio streams used to generate modulated signals.
  • Confusion matrices: At 0dB and -8dB SNR, confusion increased, particularly between QAM16 and QAM64.QAM16 is described as a subset of QAM64, contributing to their confusion.
  • Input representation: Providing normalized amplitude and phase enabled good LSTM results, whereas direct time-domain IQ input produced 9% constant accuracy on RadioML.The IQ-input models also failed to reduce training loss.
  • Interpretation: The results indicate that deep complex structures are unnecessary for good high-SNR accuracy, while convolutional layers may help at low SNRs.The low-SNR CNN comparison could not be reproduced, although its high-SNR results matched the reported results.

B. Classification accuracy on modified RadioML dataset

The LSTM was evaluated on variable-length sequences and unseen input lengths to test robustness to changing symbol rates. It retained substantial accuracy on lengths absent from training, especially at high SNR.

  • Generalization: The model’s variable-length evaluation directly tests generalization to input sequences absent from training.The model was trained on 128–512-sample inputs before evaluation on shorter inputs such as 64 samples.
  • Variable input lengths: 75% average accuracy was achieved with 64 samples and 4 samples per symbol, despite training on lengths from 128 to 512.Accuracy improved as the model saw more modulated symbols.
  • Generalization: Above 70% average accuracy was consistently achieved at high SNR when testing unseen sample lengths.Four folds containing lengths 64, 128, 256, and 512 were used, with each left-out length serving as the test fold.

C. Learned representations

Visualization and architecture studies examine how the LSTM represents modulation signals and how model capacity affects accuracy. Increasing layer depth generally improved accuracy, while additional cells provided diminishing gains at greater depth.

  • Learned representations: LSTM visualization tracks temporal cell activations and gate saturation to inspect learned representations for a QAM64 input.The analysis considers which cells activate at each time step and how long gates remain open.
  • Model capacity: The two-layer and three-layer amplitude-phase LSTM accuracy results were compared across different cell sizes on RadioML.These comparisons are presented in the corresponding cell-size accuracy figures.
  • Model capacity: Accuracy generally increased with layer depth across the tested LSTM cell sizes.Models varied from 16 to 256 cells and from one to three layers on the RadioML dataset.
  • Model capacity: Increasing the number of cells yielded limited additional improvement as model depth increased.The paper reports this pattern across the layer-depth and cell-count study.

VI. RESOURCE FRIENDLY MODELS

The paper investigates resource-friendly classification for low-cost distributed sensors by using averaged magnitude FFT data and quantized models. Sequential scanning supports wideband coverage despite limited sensor sampling rate.

  • Resource constraints: Resource-friendly models target lower data transfer, storage, and computational requirements for low-end distributed sensors.The study evaluates averaged magnitude FFT classification and quantized deep-learning models for possible sensor deployment.
  • Sensor acquisition: Electrosense sensors sample at 2.4 MS/s while sequentially tuning to centre frequencies to monitor the spectrum.Their limited sampling rate means wideband signals require sequential scans.
  • Averaged magnitude FFT: Equation 8 represents concatenating averaged magnitude FFT data from sequential scans to reconstruct the full bandwidth of interest.The concatenated centre-frequency spectra are sent to the cloud for processing.
  • Evaluation data: The averaged magnitude FFT dataset is identified as the basis for evaluating resource-friendly classification.Its parameters are listed in Table III.

B. Electrosense dataset

The Electrosense study evaluates LSTM-based technology classification from averaged magnitude FFT sequences collected by distributed sensors. The model achieves 80% accuracy, but similar power spectra create substantial confusion and limit averaged-spectrum classification.

  • Technology classification: Variable-length averaged magnitude FFT sequences are processed with the same LSTM architecture used for complex input data.The architecture is selected because it can handle variable-length inputs and learn long-term dependencies.
  • Technology classification: 80% classification accuracy is achieved on the Electrosense dataset using averaged magnitude FFT data.The data are collected over the air from multiple Electrosense sensors and fed as variable-length sequences to the LSTM.
  • Classification behavior: LTE and DVB are frequently confused because both use OFDM and have similar power spectra.This similarity illustrates why modulation classification is not always sufficient for technology classification.
  • Limitations: Averaged magnitude spectrum is sub-optimal because multiple modulation schemes can have identical power spectral densities.Examples include 8PSK, QAM16, and QPSK after pulse shaping with a fixed roll-off factor.
  • Classification behavior: 19% accuracy is obtained for all 11 RadioML modulations with magnitude FFT inputs, even at high SNRs.The modulations have identical magnitude FFTs because they share bandwidths and pulse-shaping filters.

E. Quantized models

The paper studies quantized neural-network models for low-end sensor deployment, balancing reduced resource requirements against classification accuracy. Ternary-weight LSTMs with 4-bit activations retain close to 80% accuracy, whereas binarized LSTMs perform poorly.

  • Quantized LSTM models: Close to 80% accuracy is achieved by LSTMs with ternary weights and 4-bit activations.The TW 4BA variant reduces the computational power required relative to full-precision models.
  • Quantized CNN models: Binarized CNNs achieve accuracy 10% below their full-precision variants.The paper contrasts this result with reports that binarized CNNs can approach full-precision performance.
  • Quantized LSTM models: Binarized LSTMs provide very poor accuracy, requiring the study to use 4-bit quantized LSTM variants.The paper evaluates ternary weights with either full-precision or 4-bit activations.
  • Deployment implications: Quantization can reduce resource consumption through fixed-point arithmetic, smaller intermediate storage, and lookup-table activations.These implementation benefits are described for 4-bit quantized LSTMs.
  • Deployment implications: Quantized-kernel support in standard machine-learning libraries is incomplete and was limited to a minimum of 8-bit quantization when the paper was written.The paper expects support for low-end deployment to improve as ARM-oriented kernels develop.

VII. CONCLUSION AND FUTURE WORK

The paper concludes that LSTM models can classify amplitude-phase signals effectively, accommodate variable-length inputs, and support technology classification from averaged spectra. It also identifies representation quality, low-SNR performance, generalization, and labeled-data requirements as important boundaries for future work.

  • Conclusion: LSTM models achieve state-of-the-art results at high SNRs from 0dB to 20dB using time-domain amplitude and phase inputs.The approach does not require complex CNN-LSTM models.
  • Conclusion: Variable-length LSTM inputs can capture sample-rate variations effectively.This supports classification across differing sequence lengths and sample-rate conditions.
  • Conclusion: Averaged magnitude FFT information enables basic technology classification under distributed-sensor uplink bandwidth constraints.The conclusion frames this as feasible for low-cost sensor networks such as Electrosense.
  • Conclusion: Quantized LSTMs reduce processing-power requirements at a cost of 10% accuracy loss.The paper connects this trade-off to deployment on low-cost sensor networks.
  • Future work: CNN models may provide an additional 5-10% accuracy on low-SNR conditions below -2dB.The authors state that they could not replicate the cited CNN results because of hyperparameter tuning.
  • Future work: Further analysis is required for variable symbol rates, unknown channel conditions, unknown modulation parameters, and semi-supervised learning.The paper also calls for testing models that handle all possible spread-spectrum modulations.
Loading 1707.08908v2…