[Paper Review] Deep Learning in Beyond 5G Networks with Image-based Time-Series Representation

Lucas Fernando Alvarenga e Silva, Bruno Yuji Lino Kimura+1 more|arXiv (Cornell University)|Apr 17, 2021

Wireless Signal Modulation Classification8 references4 citations

TL;DR

This paper proposes a novel deep learning approach for predicting 5G radio signal quality indicators in Beyond 5G (B5G) networks by transforming one-dimensional time-series data into image representations using techniques like Recurrence Plots, Markov Transition Fields, and Gramian Angular Fields. These images are then processed by convolutional neural networks (CNNs), achieving an average normalized root mean squared error (NRMSE) of 3.43% on 5G transmission datasets, outperforming LSTM-based models by nearly fourfold in accuracy while using significantly fewer parameters.

ABSTRACT

Towards the network innovation, the Beyond Five-Generation (B5G) networks envision the use of machine learning (ML) methods to predict the network conditions and performance indicators in order to best make decisions and allocate resources. In this paper, we propose a new ML approach to accomplish predictions in B5G networks. Instead of handling the time-series in the network domain of values, we transform them into image thus allowing to apply advanced ML methods of Computer Vision field to reach better predictions in B5G networks. Particularly, we analyze different techniques to transform time-series of network measures into image representation, e.g., Recurrence Plots, Markov Transition Fields, and Gramian Angular Fields. Then, we apply deep neural networks with convolutional layers to predict different 5G radio signal quality indicators. When comparing with other ML-based solutions, experimental results from 5G transmission datasets showed the feasibility and small prediction errors of the proposed approach.

Motivation & Objective

• To improve prediction accuracy of 5G radio signal quality indicators (e.g., CQI, SNR, RSRP) in dynamic B5G networks.
• To address the limitations of reactive network management in high-mobility, high-frequency 5G scenarios by enabling proactive decision-making through machine learning.
• To evaluate the feasibility and performance of transforming 1D network time-series into 2D image representations for enhanced deep learning-based forecasting.
• To compare the proposed image-based approach with established 1D-CNN and LSTM models in terms of prediction accuracy and computational efficiency.

Proposed method

• Transform one-dimensional 5G network time-series (CQI, SNR, RSRQ, RSRP, RSSI) into 2D image representations using three techniques: Recurrence Plots (RP), Markov Transition Fields (MTF), and Gramian Angular Fields (GAF).
• Feed the resulting images into pre-trained 2D Convolutional Neural Networks (CNNs) for regression tasks to predict future signal quality indicators.
• Use Mean Squared Error (MSE) and quantile loss functions to optimize model performance and model prediction uncertainty.
• Train and evaluate models on 83 real-world 5G datasets from a mobile operator, covering static and vehicular mobility scenarios with file download and video streaming services.
• Compare performance against 1D-CNN and LSTM baselines using the same dataset, with metrics including NRMSE and model complexity (parameters and MAC operations).
• Apply early stopping and Adam optimizer with a learning rate of 0.001 to improve training convergence and generalization.

Experimental results

Research questions

• RQ1 Can transforming 1D network time-series into 2D image representations improve prediction accuracy of 5G radio signal quality indicators in B5G networks?
• RQ2 How do different image-based time-series encoding techniques (RP, MTF, GAF) compare in their ability to preserve temporal dynamics for downstream CNN regression?
• RQ3 To what extent does the proposed image-based CNN approach outperform established 1D-CNN and LSTM models in predicting 5G signal quality under static and vehicular mobility conditions?
• RQ4 What is the trade-off between prediction accuracy and model complexity (e.g., number of parameters, MAC operations) in the proposed method versus traditional time-series models?

Key findings

• The proposed image-based deep learning approach achieved an average NRMSE of 3.43% across all test scenarios, with a standard deviation of 0.0139, indicating high consistency and low prediction error.
• The method outperformed the LSTM baseline by nearly four times in prediction accuracy, with an average NRMSE of 12.3% compared to 3.43%.
• The proposed approach achieved comparable accuracy to a 1D-CNN model (NRMSE: 2.17%) but with significantly fewer parameters—approximately 29 times fewer—making it more computationally efficient.
• The performance of different image transformation techniques (RP, MTF, GAF) was highly similar, with MTF-based CNN achieving the best average NRMSE of 3.43%.
• Prediction errors were consistently below 10% for long-range indicators (SNR, RSRP, RSRQ, RSSI), indicating robustness across diverse signal quality metrics.
• The method demonstrated strong generalization across diverse mobility patterns (static and vehicular) and application types (file download and video streaming), confirming its real-world applicability.