A Big Data Enabled Channel Model for 5G Wireless Communication Systems

1 INTRODUCTION

• 5G’s larger bandwidths, massive antenna arrays, dense deployments, and varied scenarios are expected to generate large wireless datasets.The introduction also describes rapid mobile-data growth and evolving devices as contributors.Source: §1 · p. 1
• Wireless big data is multi-dimensional, personalized, multi-sensory, and real-time, with correlations across time, location, and social relationships.These properties are described as additional to the commonly cited five V’s of big data.Source: §1 · pp. 1–2
• Channel statistics such as received power, RMS delay spread, and RMS angle spread can have complicated nonlinear relationships with network layout.The relevant layout factors include Tx/Rx locations, carrier frequency, and scatterer distributions.Source: §1 · p. 2
• The paper proposes FNN- and RBF-NN-based channel models learned from datasets, reporting a trade-off among accuracy, complexity, and flexibility.The authors contrast direct learning from datasets with existing models that rely on assumptions or repeated configuration for different layouts.Source: §1 · p. 2

2 OVERVIEW OF BIG DATA ANALYTICS IN WIRELESS COMMUNICATIONS

• Big data analytics includes stochastic modeling, data mining, and machine learning; the overview focuses on machine learning’s wireless and channel-modeling applications.It distinguishes supervised learning, which maps inputs to outputs, unsupervised learning, which models input structure, and reinforcement learning, which uses reward feedback.Source: §2 · p. 2, §2.1 · p. 2

2.2 Machine Learning in Wireless Communications

• Wireless communication problems can be framed as clustering, classification, and regression, with learned datasets used to support adaptation across applications.Examples include localization, NLOS identification and mitigation, spectrum sensing, antenna selection, caching, and resource allocation.Source: §2.2 · pp. 2–3

2.3 Machine Learning in Wireless Channel Modeling

• Emerging 5G scenarios and channel characteristics call for large channel-response datasets and measurements spanning new propagation conditions.Examples include 3-D double-directional angles, non-stationarity, spherical wavefronts, high path loss, and high delay resolution.Source: §2.3 · p. 3
• Prior channel-modeling work applies learning to path-loss prediction, direction estimation, MPC processing, CIR modeling, and wireless-channel classification.The cited approaches include RBF-NN, RVM, neural networks, CNN, SVM, clustering, and complex LASSO.Source: §2.3 · pp. 3–4

3 A MACHINE LEARNING ENABLED CHANNEL MODEL

• Broad measurement datasets are difficult to obtain because 5G-capable channel sounders and measurement campaigns are expensive and time-consuming.The measurement datasets used here come from fixed locations, while simulation datasets are generated randomly.Source: §3 · p. 4

3.1 Channel Measurement Datasets

• The indoor measurement campaign covered four mmWave bands—11, 16, 28, and 38 GHz—with four Tx locations and a fixed Rx location.The room measured 7.2×7.2×3 m3; the measurement data were collected using scanned arrays and processed with SAGE to extract MPC parameters.Source: §3.1 · p. 4, Figure 1 · p. 4

3.2 GBSM Simulation Datasets

• The GBSM represents scatterers with geometric shapes and derives MPC power, delay, and departure and arrival angles from geometric relationships.The modeled inputs include Tx/Rx coordinates, Tx–Rx distance, and carrier frequency.Source: §3.2 · pp. 4–5

3.3 ANN Based Channel Model

• The framework compares FNN and RBF-NN; RBF-NN adds neurons during learning and adjusts one layer of weights, reducing learning cost relative to FNN.The input parameters are normalized to the range −1 to 1 before training.Source: §3.3 · p. 5
• The ANN maps eight Tx/Rx and frequency-related inputs to six channel statistics, including received power, RMS delay spread, and four angle spreads.The training and testing data combine measurements and GBSM simulations; the simulation inputs vary randomly in a confined environment and across 10–60 GHz.Source: §3.3 · p. 5

4 RESULTS AND ANALYSIS

• Testing hidden-layer sizes identified 10 hidden neurons as achieving good performance with relatively low complexity; the resulting FNN uses an 8-10-6 architecture and convergence-error goal 5 × 10−2.Source: §4 · p. 6
• The RBF-NN uses a Gaussian basis with spread σ = 0.75 for a smooth interpolation plane and a convergence-error goal of 3 × 10−2.Five neurons are added between two displays during training.Source: §4 · p. 6
• Both networks are trained with Levenberg-Marquardt for 1,000 iterations and evaluated using mean square error; the FNN goal is larger to avoid local optimization and over-fitting.Table 3 summarizes the detailed FNN and RBF-NN parameters.Source: §4 · p. 6, Table 3 · p. 6
• After training, the networks predict outputs for testing inputs, which are compared with measured and simulated outputs to evaluate performance.Figure 2 presents the general ANN-based channel-model framework.Source: §4 · p. 6, Figure 2 · p. 6

4.1 Predicting Statistical Properties for Measurement Datasets

• Both FNN and RBF-NN work well on measurement datasets, accurately predicting received power, RMS DS, and RMS AAS, while RMS EAS has slightly larger prediction error.The measurement and predicted values are shown for received power, RMS DS, RMS AAS, and RMS EAS.Source: §4.1 · p. 6
• The measurement-dataset predictions cover received power, RMS DS, RMS AAS, and RMS EAS.Source: §4.1 · p. 6

4.2 Predicting Statistical Properties for Simulation Datasets

• The simulated outputs span −80 dB to −60 dB for received power, 2 ns to 12 ns for RMS DS, and 20° to 140° azimuth and 10° to 50° elevation angles.The GBSM simulation datasets have statistical properties similar to the channel measurement datasets.Source: §4.2 · pp. 6–7
• Predicted values and trends generally match the simulated outputs, though extreme real values can have larger deviations.The figures cover received power, RMS DS, and RMS angle spreads.Source: §4.2 · pp. 6–7, Figure 7 · p. 7, Figure 8 · p. 7
• Received power and RMS DS predictions are better than RMS AS predictions, possibly because the former properties correlate more strongly with the inputs.The authors relate received power to Tx–Rx distance and RMS AS to scatterer-dependent multipath angles.Source: §4.2 · p. 7
• RBF-NN generally performs better than FNN, whose outputs can vary across runs because of initialization and local optimization.FNN may also fail to reach its convergence-error goal.Source: §4.2 · p. 7
• The framework extends beyond received power or path loss to include delay and 3-D double-directional angular channel sparsity properties.Its inputs include accurate horizontal Tx and Rx locations and carrier frequencies spanning bands including 11, 16, 28, 38, 45, and 60 GHz.Source: §4.2 · p. 7

4.3 Prediction Performance Comparison

• A convergence-error goal near zero can overfit training data, with extreme unknown inputs producing large deviations or unreasonable predictions.The simulation convergence goals were tuned for good performance.Source: §4.3 · p. 7
• RMSE comparisons show better prediction of received power and RMS DS than RMS AS, and better performance for RBF-NN than FNN.Table 4 compares RMSEs for total, training, and testing datasets.Source: §4.3 · pp. 7–8, Table 4 · p. 8
• Total-dataset and testing-dataset RMSEs are a little larger than training-dataset RMSEs.The comparison covers received power, RMS DS, and RMS AS properties.Source: §4.3 · p. 8

4.4 Input and Output Relationships

• The example varies carrier frequency from 10 GHz to 40 GHz and the Rx y-coordinate from 0.6 m to 6.6 m, while other antenna coordinates are fixed.The Tx antenna is fixed at (1,3,1.45), and the Rx x- and z-coordinates are 4 m and 2.6 m.Source: §4.4 · p. 8
• FNN and RBF-NN predict similar received-power trends across the varying inputs, while RBF-NN surfaces are smoother.Source: §4.4 · p. 8
• Predicted received power tends to decrease as carrier frequency increases; varying the Tx y-coordinate also varies Tx–Rx distance and received power.Source: §4.4 · p. 8

4.5 Extensions and Discussions

• FNN and RBF-NN perform similarly on measurement datasets, while RBF-NN performs better on simulation datasets; real measurements better reflect environmental objects.The authors also report that measured data can capture more complex input-output relationships.Source: §4.5 · p. 8, §4.5.2 · p. 9

4.5.1 Scenario Representation

• A general representation of environmental objects across settings is needed for multi-environment datasets, whose channel-statistic fingerprints can support scenario classification.The discussion names indoor objects such as tables, desks, and electronic devices, and outdoor objects such as trees, buildings, and cars.Source: §4.5.1 · p. 9

4.5.2 Big Datasets Acquisition

• Because measurements across environments are expensive and time-consuming, open-access channel datasets are proposed to accelerate research; WiCO reports measurements across several typical scenarios.The listed scenarios include beach, stadium, hotel lobby, rural area, and campus hot spots.Source: §4.5.2 · p. 9

4.5.3 Inputs and Outputs Selection

• Received power and RMS DS correlate more strongly than RMS ASs with Tx/Rx coordinates, Tx–Rx distance, and carrier frequency, motivating selection of inputs most correlated with outputs.Source: §4.5.3 · p. 9

• The paper proposes an ANN-based channel-model framework using datasets collected from real channel measurements and GBSM simulations.Source: §5 · p. 9
• FNN and RBF-NN predict received power, RMS DS, and RMS ASs in the proposed framework.Source: §5 · p. 9
• The analyzed and validated simulation results indicate that machine learning can be a powerful analytical tool for future measurement-based wireless channel modeling.Source: §5 · p. 9