Lightweight CFR-Based Modulation Adaptation in a Real-Time MIMO-OFDM SDR Testbed

Motivation

• Scalar SNR-based adaptation is efficient but does not retain the frequency-selective structure of MIMO-OFDM channels.Different channel realisations can have similar average SNR while differing across subcarriers and antenna pairs.Source: §I · p. 1
• Full complex CSI can improve link adaptation but increases learning dimensionality and implementation complexity.Prior work using a full complex channel matrix with CNN-LSTM achieved 92.5% test accuracy.Source: A. Related Work · p. 1

Approach

• CFR provides an intermediate representation by compressing frequency-dependent channel information into mean, standard deviation, range, and slope descriptors.The descriptors represent average channel magnitude and its variation across frequency.Source: A. Related Work · p. 2

Evaluation

• The study compares BS-side SNR, CFR descriptors, and their combination using DT, RF, and KNN classifiers on OTA measurements near modulation reliability boundaries.Modulation adaptation is evaluated independently of coding-rate optimisation.Source: A. Related Work · p. 2

Results

• CFR-based classifiers outperform SNR-only classifiers, reaching 73.6%–81.4% test accuracy versus 35%–42%.Performance is maintained or improved near the 10% BLER reliability thresholds, reaching 82.8% for the RF CFR+SNR classifier.Source: A. Related Work · p. 2

Implementation

• A depth-7 DT with 123 leaves is selected for real-time deployment because its inference can be implemented as compact conditional C code.The classifier was integrated into the LabVIEW C Node without online retraining or GPU-based inference.Source: A. Related Work · p. 2

II. SYSTEM MODEL

• The section defines the system overview, signal model, and OTA measurement configuration for the study.The detailed hardware and measurement conditions are described in subsequent subsections.Source: §II · p. 2

System Overview

• The real-time testbed uses NI USRP-2954 SDR hardware and LabVIEW Communications in a 14-BS-antenna, 1-MS-antenna configuration.The resulting frequency-dependent channel matrix is H[k] ∈ C^Nr×N_t.Source: A. System Overview · p. 2

Modulation Selection

• Under channel reciprocity, the uplink-derived CFR is used to extract magnitude descriptors for the ML-based modulation selector.The selected modulation index is passed to the downlink frame scheduler and OFDM modulator.Source: A. System Overview · p. 2, B. MIMO-OFDM Signal Model · p. 2

Channel Acquisition

• Uplink pilots transmitted by the MS are processed at the BS to estimate the uplink CFR using a least-squares algorithm.The BS uses this estimate as the corresponding downlink channel representation under TDD reciprocity.Source: B. MIMO-OFDM Signal Model · pp. 2–3

Measurement Roles

• The architecture provides CFR-derived features to the deployed decision-tree classifier, while MS-side SNR and BLER are reserved for measurement analysis and evaluation.This separates BS-side inference inputs from MS-side performance measurements.Source: Figure 1 · p. 3, B. MIMO-OFDM Signal Model · p. 3

OTA Measurement Setup

• Measurements are collected at 4.2 GHz over a 20 MHz channel bandwidth with the MS held stationary.The stationary configuration supports repeatable channel conditions during data acquisition.Source: C. OTA Measurement Setup · p. 3
• The OTA setup uses an NI USRP-2954 testbed with 14 BS antennas and 1 MS antenna separated by 60 cm.The configuration creates a line-of-sight propagation scenario.Source: Figure 2 · p. 3, C. OTA Measurement Setup · p. 3

• The testbed sweeps BS transmit power across 100 levels for each modulation and records measurements during fixed 30 s intervals.Source: A. Data acquisition · p. 3
• BS-side CFR and SNRBS are sampled at 10 Hz, while MS-side SNRMS and BLERMS are recorded at 1 Hz and aligned by timestamp.Source: A. Data acquisition · p. 3
• BS-side measurements supply classifier inputs, whereas MS-side measurements support BLER fitting, threshold derivation, and offline analysis.Source: A. Data acquisition · p. 3
• CFR magnitude is smoothed before extracting features that represent average channel quality and frequency-selective variation.Source: A. Data acquisition · p. 3
• A logistic sigmoid is fitted to each modulation’s BLERMS–SNRMS relationship, and the 10% BLER SNR defines reliability and transition regions.Source: A. Data acquisition · pp. 3–4

• Four CFR descriptors capture mean magnitude, standard deviation, range, and first-order least-squares slope across subcarriers.Source: B. Feature Engineering · p. 4
• The CFR descriptors are combined with BS-side SNR to form candidate feature vectors for modulation classification.Source: B. Feature Engineering · p. 4
• Three feature configurations are evaluated to isolate the contribution of channel representation from classifier architecture.Source: B. Feature Engineering · p. 4
• MS-side SNRMS and BLERMS are excluded from classifier inputs and reserved for threshold derivation and offline performance analysis.Source: B. Feature Engineering · p. 4
• Feature importance analysis tests whether performance is associated mainly with average CFR magnitude or its variation across frequency.Source: B. Feature Engineering · p. 4

• QPSK, 16-QAM, and 64-QAM are used as the three output classes, with transmitted modulation serving as the training label.Source: C. Machine Learning-Based Modulation… · p. 4
• DT, RF, and KNN provide complementary comparisons of interpretable partitioning, ensemble robustness, and non-parametric classification.Source: C. Machine Learning-Based Modulation… · p. 4
• The dataset uses an 80% training and 20% test split, with 5-fold cross-validation for model selection and tuning.Source: C. Machine Learning-Based Modulation… · p. 4
• The classifiers are evaluated with SNR-only, CFR-only, and CFR+SNR BS-side feature configurations.Source: C. Machine Learning-Based Modulation… · p. 4
• Real-time inference runs at 10 Hz, and modulation updates require three consecutive identical predictions, corresponding to 0.30 s.Source: C. Machine Learning-Based Modulation… · p. 4

• The final dataset contains 87,817 OTA observations after constraint-mask filtering.Source: A. Data Acquisition & Channel… · p. 4
• The measured BLERMS–SNRMS relationship is summarized with fitted sigmoid curves and 10% BLER reliability thresholds.Source: A. Data Acquisition & Channel… · pp. 4–5
• The thresholds are 13.0 dB, 20.4 dB, and 27.0 dB for QPSK, 16-QAM, and 64-QAM, respectively.Source: A. Data Acquisition & Channel… · p. 5
• Only 0.1% of QPSK samples fall below its threshold, while 5.3% of 64-QAM samples fall above its threshold.Source: A. Data Acquisition & Channel… · p. 5
• BS-side SNR class means cluster at 28.2, 28.4, and 28.9 dB with within-class standard deviations of 1.7–2.3 dB, producing strong overlap.Source: A. Data Acquisition & Channel… · p. 5

• CFR-derived features substantially improve classification accuracy compared with SNR-only inputs across DT, RF, and KNN.Source: B. Offline Classification Performance · p. 5
• Adding SNRBS to CFR features provides little additional benefit, indicating that CFR captures most useful discriminative information under the considered conditions.Source: B. Offline Classification Performance · p. 5
• RF achieves the highest offline accuracy among the evaluated classifiers.Source: B. Offline Classification Performance · p. 5
• A depth-7 DT with 123 leaves is selected by 5-fold cross-validation as a suitable performance–complexity operating point for real-time deployment.Source: B. Offline Classification Performance · p. 5

• The CFR+SNR feature set is analysed with DT and RF to identify which BS-side measurements drive classification improvement.Source: C. Feature Importance · p. 5
• SNRBS contributes only 1.0% for DT and 9.5% for RF, substantially less than the CFR-derived features.Source: C. Feature Importance · p. 5, Figure 5 · p. 6
• Mean CFR magnitude |H[k]| is the most influential feature, accounting for 63.3% of DT importance and 38.0% of RF importance.Source: C. Feature Importance · p. 5
• The remaining importance is distributed across CFR range, standard deviation, and slope, capturing both average magnitude and frequency-domain variation.Source: C. Feature Importance · p. 5
• Near the 10% BLER thresholds, CFR-based models maintain or improve performance, while adding SNRBS again provides limited additional benefit.Source: D. Transition Region Analysis · p. 6

• CFR-based features consistently outperform the SNR-only baseline across 87,817 OTA observations, with little additional benefit from adding SNRBS.Source: §V · p. 6
• CFR-based performance is retained near the measured 10% BLER reliability thresholds, and mean CFR magnitude is the dominant input.Source: §V · p. 6
• A depth-7 DT with 123 leaves is integrated into the LabVIEW C Node for real-time inference at 10 Hz as a practical performance–complexity trade-off.Source: §V · p. 6
• The evaluation is limited to stationary indoor LoS conditions and magnitude-only CFR features.Source: §V · p. 6