Big Data Analysis and Prediction of Electromagnetic Spectrum Resources: A Graph Approach

Han Zhang, Siqi Peng, Jingyu Zhang, Yun LinSustainability 2022 · DOI

TL;DR

This paper tackles the growing challenge of scarce and complex electromagnetic spectrum usage by proposing a novel graph-based deep learning architecture. By modeling spatial correlations in spectrum activity, their approach significantly improves prediction accuracy, especially in large-scale, complex environments, offering a more sustainable approach to spectrum management.

What Problem Does This Paper Address?

The wireless communication industry is booming, leading to an explosion of connected devices. This surge strains limited spectrum resources, making them scarcer and the electromagnetic environment increasingly complex. This complexity can lead to interference and frequency conflicts, threatening the sustainable development of these industries. Existing methods often struggle to effectively analyze and predict spectrum activity, particularly when dealing with the intricate spatial relationships of spectrum usage. The paper aims to bridge this gap by developing a method that can mine and analyze spectrum activity rules to mitigate frequency conflicts and ensure efficient spectrum utilization.

Methods

The core of this paper is a novel graph-based joint analysis and prediction architecture. Instead of treating spectrum data as purely temporal or frequency-based, it incorporates spatial relationships.

  1. Spatial Correlation Modeling: The approach begins by extracting various features (duty cycle, mean value, standard deviation, quartile distance, autocorrelation coefficient, entropy, skewness, and kurtosis) from spectrum activity data at different locations. These features are then used to construct a feature matrix for each spatial location. A spatial correlation coefficient is calculated between these matrices to quantify the similarity of spectrum activity between different locations.
  2. Graph Construction: Based on these correlation coefficients, an adjacency matrix is built. This matrix represents the graph structure, where nodes are spatial locations and edges signify the strength of their spectral activity correlation.
  3. Graph Convolutional Network (GCN) Architecture: A deep learning model is designed, incorporating Graph Convolutional Networks (GCNs) and dilated convolutions.
  4. GCNs are used to learn spatial dependencies by leveraging the adjacency matrix, effectively capturing the non-linear relationships in spatial data that traditional convolutional methods struggle with.
  5. Dilated convolutions are employed to capture time-frequency features, allowing the model to learn from long-term historical data without losing resolution, which is crucial for predicting future spectrum usage.
  6. The architecture consists of stacked "Time-Frequency-Space" (TFS) blocks, each containing both dilated convolution and graph convolution units, connected via skip connections to enhance information flow.

Datasets

The authors evaluated their proposed method using two real-world datasets derived from the open-source ElectroSense project:

  • Madrid Dataset: A small-scale dataset collected from multiple sensors in Madrid, Spain, covering an area of approximately 15 km x 10 km over 14 days.
  • Switzerland Dataset: A large-scale dataset comprising data from multiple sensors across several cities in Switzerland (Zurich, Bern, Lausanne, St. Gallen), covering an area of approximately 120 km x 210 km over the same 14-day period.

Both datasets focus on the 790-820 MHz frequency band with a time resolution of 15 minutes and a frequency resolution of 1 MHz. Missing data in both datasets were handled through linear interpolation.

Key Insights and Ideas

  • Spatial correlations are crucial: The paper strongly advocates for incorporating spatial relationships into spectrum analysis and prediction, moving beyond purely temporal or frequency-based models.
  • Graphs are a natural fit for spatial spectrum data: The non-linear and irregular distribution of spectrum activity across space makes graph-based approaches particularly well-suited, outperforming traditional convolutional methods in complex spatial scenarios.
  • Multi-domain integration is the future: The results suggest that combining time, frequency, and spatial information leads to superior prediction accuracy compared to single-domain approaches.
  • Prior information enhances prediction: By using the adjacency matrix derived from spatial correlations as prior information, the GCN can learn spatial dependencies more effectively, leading to improved prediction accuracy.

Key Results

The proposed graph-based approach demonstrated superior performance compared to several baseline models (ARIMA, 1D LSTM, 2D LSTM, CNN+LSTM, ConvLSTM) on both datasets.

  • Madrid Dataset (small scale): The proposed method achieved a Mean Absolute Error (MAE) of 0.9094, which was 8.3% lower than the next best model (ConvLSTM).
  • Switzerland Dataset (large scale): The proposed method achieved an MAE of 0.8969, which was 11.7% lower than the CNN+LSTM model.
  • The paper highlights that the performance advantage of the proposed method becomes more pronounced with increasing prediction steps, indicating its robustness for longer-term forecasting.
  • The results also show that multi-domain joint prediction models generally outperform single time-series prediction models.

Limitations and Future Work

The authors acknowledge several areas for improvement and future research:

  • Feature Selection Optimization: The current feature selection is based on experience and assigns equal weight to all features. Future work could explore advanced optimization algorithms (e.g., heuristic, meta-heuristic) to dynamically select and weight features for better spectrum analysis and prediction. A challenge here is the lack of a universally accepted metric for measuring analysis result quality beyond final prediction accuracy.
  • Online Learning and Deployment: The current model is trained offline. For practical deployment, where spectrum usage is constantly changing, online learning strategies and transfer learning are proposed to enable models to adapt to new data and learn continuously, speeding up deployment.

Summary updated

Abstract

In the field of wireless communication, the increasing number of devices makes limited spectrum resources more scarce and accelerates the complexity of the electromagnetic environment, posing a serious threat to the sustainability of the industry’s development. Therefore, new effective technical methods are needed to mine and analyze the activity rules of spectrum resources to reduce the risk of frequency conflict. This paper introduces the idea of graphs and proposes a spectrum resource analysis and prediction architecture based on big data. In this architecture, a spatial correlation model of spectrum activities is constructed through feature extraction. In addition, based on this correlation model, a depth learning network based on graph convolution is designed, which uses the prior information of spatial activity to achieve the efficient prediction of spectrum resources. Numerical experiments were carried out on two datasets with different spatial scales. Compared with the best baseline model, the prediction error is reduced by 8.3% on the small-scale dataset and 11.7% on the large-scale dataset. This shows that the proposed method is applicable to different spatial scales and has more obvious advantages in complex scenes with large spatial scales. It can effectively use the results of spatial domain analysis to improve the prediction accuracy of spectrum resources.
BibTeX
@inproceedings{zhang2022big,
  author = {Han Zhang and Siqi Peng and Jingyu Zhang and Yun Lin},
  title = {{Big Data Analysis and Prediction of Electromagnetic Spectrum Resources: A Graph Approach}},
  booktitle = {Sustainability},
  year = {2022},
  doi = {10.3390/su15010508},
}
Figure