Signal Demodulation with Machine Learning Methods for Physical Layer Visible Light Communications: Prototype Platform, Open Dataset and Algorithms

Shuai Ma, Jiahui Dai, Songtao Lu, Hang Li, Han Zhang, Chun Du, Shiyin LiIEEE Access 2019 · DOI · OpenReview

TL;DR

This paper introduces a novel approach to demodulating signals in Visible Light Communication (VLC) systems using machine learning. The authors developed a flexible VLC prototype, collected a unique real-world dataset of modulated signals, and proposed three ML-based demodulators (CNN, DBN, AdaBoost), demonstrating AdaBoost's superior performance in this challenging communication environment.

What Problem Does This Paper Address?

Visible Light Communication (VLC) offers a promising alternative to traditional radio frequency (RF) systems, especially in crowded spectrum environments. However, demodulating signals in VLC presents unique challenges. Unlike RF, VLC signals must be real and non-negative due to the intensity modulation and direct detection (IM/DD) mechanism. Furthermore, practical VLC channels are complex and influenced by numerous factors like LED bandwidth limitations, multipath dispersion, and noise, making traditional model-based demodulators less effective. Existing research often relies on synthetic data or makes simplifying assumptions about the channel. This paper aims to address the need for robust, data-driven demodulation methods for VLC systems using real-world data.

Methods

The authors propose three distinct machine learning-based demodulators:

  • Convolutional Neural Network (CNN): This approach treats the received signal as an image. The signal is first transformed into a 2D waveform image, then resized and binarized. This image is then fed into a CNN with two convolutional layers and two pooling layers for feature extraction and classification.
  • Deep Belief Network (DBN): This method utilizes a DBN composed of three Restricted Boltzmann Machines (RBMs) to extract hierarchical modulation features from the signal. The DBN is pre-trained using gradient descent and then fine-tuned with backpropagation.
  • Adaptive Boosting (AdaBoost): This technique employs the AdaBoost algorithm to build a strong classifier by combining multiple weak classifiers. The paper specifically uses k-Nearest Neighbors (KNN) as the weak classifier. AdaBoost iteratively trains KNN classifiers, adjusting sample weights to focus on misclassified samples.

A key aspect of their methodology is the development of a flexible end-to-end VLC system prototype to collect real-world data.

Datasets

The paper introduces a novel, publicly available real modulated signal dataset for VLC systems. This dataset was collected using their prototype and comprises eight types of modulated signals: OOK, QPSK, 4-PPM, 16-QAM, 32-QAM, 64-QAM, 128-QAM, and 256-QAM. For each modulation type, data was collected at various transmission distances (0cm to 140cm) and with different numbers of sample points per period (N=10, 20, 40, 80). This dataset is a significant contribution, as it provides a realistic benchmark for evaluating VLC demodulation techniques.

Key Insights and Ideas

  • Data-driven demodulation is crucial for complex VLC channels: The paper strongly advocates for ML-based, model-free demodulators due to the inherent complexities and unpredictability of practical VLC channels, which are difficult to model accurately.
  • Signal-to-image transformation for CNNs: Representing the time-series signal data as an image allows CNNs, traditionally used for image recognition, to be effectively applied to signal demodulation.
  • AdaBoost excels in this domain: The experimental results consistently show AdaBoost outperforming CNN and DBN, suggesting its robustness in handling the variations and noise present in VLC signals.
  • Modulation order vs. distance trade-off: Higher modulation orders (e.g., 256-QAM) offer higher data rates but are more susceptible to distance and noise, leading to lower demodulation accuracy. For short distances or high SNR, higher-order modulation is beneficial, but for longer distances, lower-order modulation becomes more reliable.

Key Results

  • AdaBoost achieves the best performance: Across various scenarios, the AdaBoost-based demodulator consistently demonstrated higher accuracy compared to CNN and DBN.
  • Demodulation accuracy degrades with distance: As the transmission distance increases, the demodulation accuracy for all proposed ML methods decreases.
  • Higher modulation order reduces accuracy: For a fixed transmission distance, increasing the modulation order negatively impacts demodulation accuracy.
  • Sample points per period (N) impact accuracy: Increasing the number of sample points per period (N) generally improves demodulation accuracy, though with diminishing returns.
  • Training data size matters: Demodulation accuracy increases with the size of the training dataset, eventually plateauing.

Limitations and Future Work

  • CNN performance limitations: The paper notes that CNNs might ignore the classical sampling theorem, potentially limiting their guaranteed performance. The pooling operation in CNNs can also lead to the loss of relative position information within waveforms.
  • MLD classification degradation: The Maximum Likelihood (MLD) classification method showed degraded performance in practical VLC channels due to their complex interference characteristics.
  • Future work: The authors express their intention to further investigate dedicated ML-based demodulators for VLC systems, suggesting ongoing research in this area.

Summary updated

Abstract

In this paper, we investigate the design and implementation of machine learning (ML) based demodulation methods in the physical layer of visible light communication (VLC) systems. We build a flexible hardware prototype of an end-to-end VLC system, from which the received signals are collected as the real data. The dataset is available online, which contains eight types of modulated signals. Then, we propose three ML demodulators based on convolutional neural network (CNN), deep belief network (DBN), and adaptive boosting (AdaBoost), respectively. Specifically, the CNN based demodulator converts the modulated signals to images and recognizes the signals by the image classification. The proposed DBN based demodulator contains three restricted Boltzmann machines (RBMs) to extract the modulation features. The AdaBoost method includes a strong classifier that is constructed by the weak classifiers with the k-nearest neighbor (KNN) algorithm. These three demodulators are trained and tested by our online open dataset. Experimental results show that the demodulation accuracy of the three data-driven demodulators drops as the transmission distance increases. A higher modulation order negatively influences the accuracy for a given transmission distance. Among the three ML methods, the AdaBoost modulator achieves the best performance.
BibTeX
@inproceedings{ma2019signal,
  author = {Shuai Ma and Jiahui Dai and Songtao Lu and Hang Li and Han Zhang and Chun Du and Shiyin Li},
  title = {{Signal Demodulation with Machine Learning Methods for Physical Layer Visible Light Communications: Prototype Platform, Open Dataset and Algorithms}},
  booktitle = {IEEE Access},
  year = {2019},
  doi = {10.1109/access.2019.2903375},
}
Figure