PaperlayerView exact revision on arXiv

Source-linked AI summary

On the Impact of Site-Specific Training for a Real-World 5G NR System

Reinhard Wiesmayr, Nuri Berke Baytekin, Chris Dick, Christoph Studer

arXiv:2609.04004v1cs.ITeess.SPmath.IT

TL;DR

Real-world evidence for site-specific receiver finetuning has largely focused on fully trainable neural receivers and single-layer transmission. This paper evaluates three receiver architectures on measured 5G NR data and finds broad finetuning gains, while site-specific covariance estimation with iterative detection achieves the study’s lowest dataset BLER.

  • Problem

    Real-world site-specific finetuning evidence has focused on fully trainable neural receivers and single-layer transmission, leaving other architectures, dual-layer transmission, and temporal robustness to assess.

    Source: A. Contributions · p. 1

  • Method

    The study evaluates three receiver architectures on over-the-air 5G NR measurements and compares site-specific finetuning with LMMSE channel estimation using synthetic- or measurement-derived covariance matrices.

    Source: A. Contributions · p. 1

  • Results

    Finetuning substantially improves NRX and MDX but yields marginal gains for DUIDD; site-specific covariance estimation with IDD achieves the lowest dataset BLER reported in the study.

    Source: §VII · p. 8

  • Takeaways & Limitations

    One NRX finetuned on single- and dual-layer data performs close to separately finetuned receivers, and finetuning remains effective across campaigns more than half a year apart.

    Source: §VII · p. 8

Abstract

from arXiv · show

Site-specific training can improve wireless receiver performance without increasing computational complexity. However, real-world results have so far focused on fully trainable neural receivers and single-layer transmissions. We study site-specific finetuning of three receiver architectures: fully trainable neural, model-driven neural, and model-based. We train and evaluate these receivers using new measurements from a standard-compliant 5G NR testbed at ETH Zurich with dual-layer uplink transmission, including measurement campaigns conducted more than six months apart. Our results show that site-specific finetuning (i) substantially improves fully trainable and model-driven neural receivers, while resulting in only marginal gains for the less tunable model-based receiver; (ii) enables a single neural receiver jointly finetuned for single- and dual-layer transmission to closely match receivers finetuned separately for each configuration; and (iii) remains effective across measurement campaigns separated by more than six months. We also investigate site-specific linear minimum mean-square error channel estimation using covariance matrices estimated from either synthetic channels or site-specific measurements. When combined with iterative detection and decoding, site-specific channel estimation achieves the lowest error rate observed in our datasets. Our finetuning code and measurement datasets are publicly available at https://github.com/IIP-Group/site_specific_training

I. INTRODUCTION

Site-specific receiver adaptation addresses the gap between synthetic training results and limited real-world evidence, extending study to multiple architectures, transmission layers, and measurement campaigns. The paper also compares neural finetuning with statistical channel-estimation adaptation.

I. INTRODUCTION

  • Site-specific finetuning uses labeled data and stochastic gradient descent, whereas parameter adaptation estimates classical receiver parameters such as channel statistics for a given site.The studied architectures span fully tunable neural, model-driven neural, and model-based receivers.Source: §I · p. 1
  • Earlier real-world finetuning evidence covered one receiver architecture and single-user, single-layer transmission, despite reported gains on held-out users.Prior real-world evaluations also used a custom single-antenna SDR setup or successful-transmission labels only.Source: §I · p. 1, B. Related Work · p. 2

A. Contributions

  • The study evaluates finetuning across NRX, MDX, and DUIDD, including dual-layer transmission and measurement campaigns more than half a year apart, and compares site-specific channel-statistics adaptation.The measurement datasets and finetuning code are publicly available.Source: A. Contributions · p. 1

B. Related Work

  • The paper estimates site-specific covariance matrices from noisy least-squares observations with noise-statistics correction and compares them with synthetic-channel covariance estimates.Prior measured-data channel-estimation work cited here used noisy channel estimates as labels.Source: B. Related Work · p. 2

II. TUNABLE RECEIVER ARCHITECTURES

  • The finetuning experiments use receivers with least-squares channel estimates; more powerful LMMSE channel estimation is examined separately.Source: §II · p. 2

A. Fully-Tunable Neural Receiver (NRX)

The paper evaluates three tunable receiver architectures alongside two indoor deployment scenarios on a standard-compliant 5G NR testbed. The receivers differ in their neural or model-based design and in how they use channel estimates and decoding.

A. Fully-Tunable Neural Receiver (NRX)

  • NRX uses a convolutional and graph neural network with eight unrolled iterations; evaluation can stop after two iterations to form the Shallow NRX.The Shallow and Deep NRX contain approximately 1.2 · 10^5 and 4.4 · 10^5 active parameters, respectively.Source: A. Fully-Tunable Neural Receiver (NRX) · p. 2

B. Model-Driven Neural Receiver (MDX)

  • MDX combines conventional receiver algorithms with a small neural network for joint channel estimation and data detection.A ResNet combines pilot-aided and data-aided channel estimates before final LMMSE maxlog soft-output detection.Source: B. Model-Driven Neural Receiver (MDX) · p. 2

C. Model-Based DUIDD Receiver

  • DUIDD deep-unfolds iterative detection and decoding, tuning soft-information exchange, decoder-message damping, and decoder-state forwarding.The evaluated DUIDD has 30 tunable parameters and uses two MMSE-PIC detection stages with 12 total decoder iterations.Source: C. Model-Based DUIDD Receiver · p. 2

III. SITE-SPECIFIC DEPLOYMENT SCENARIOS

  • Measurements use commercial 5G NR equipment in two indoor scenarios, with four-antenna O-RUs and low-SNR 16-QAM operating points in the reference receiver’s BLER waterfall region.The June 2026 figures show floor plans and serving O-RUs; the resource-grid figure depicts the first-layer DMRSs and second-layer DMRSs on even subcarriers at symbols 1, 6, and 11.Source: §III · pp. 2–3, Figure 1 · p. 3, Figure 2 · p. 3

A. Indoor Small Laboratory Scenario

The datasets cover two indoor deployments and campaigns in November 2025 and June 2026. The later campaigns add dual-layer transmission and were conducted more than half a year after the earlier ones.

1) Dataset from Nov. 2025 with Single-Layer Transmission:

  • The November 2025 small-laboratory campaign covered about 3.5 m × 3.5 m and included single-layer measurements with a Galaxy S23 and an iPhone 14 Pro at a 7 dB PUSCH SNR target.Source: 1) Dataset from Nov. 2025 with… · p. 3
  • In June 2026, the small-laboratory campaign covered 5 m × 4 m with one serving O-RU and included single-layer Galaxy S23 and dual-layer Pixel 9 Pro measurements.Their PUSCH SNR targets were 5 dB and 12 dB, respectively; the campaign followed the November campaign by more than half a year.Source: 1) Dataset from Nov. 2025 with… · p. 3

B. Indoor Large Office Floor Scenario

  • The November 2025 large-office campaign covered about 50 m × 30 m, including corridors and adjacent rooms with both LOS and NLOS conditions.It included two single-layer measurements with the same UEs and SNR targets as the November small-laboratory campaign.Source: B. Indoor Large Office Floor Scenario · p. 3
  • The June 2026 large-office campaign covered a corridor area of approximately 16 m × 12.5 m and used the same UEs, layer configurations, SNR targets, and DMRS configurations as the June small-laboratory campaign.It took place in the same large-office environment more than half a year after the November campaign.Source: B. Indoor Large Office Floor Scenario · p. 3

IV. EVALUATION PROTOCOL

The evaluation pipeline extracts labeled slot samples, creates disjoint finetuning and test sets, and finetunes pretrained receivers. Comparisons use common test samples within each receiver comparison.

A. Finetuning Data Extraction Pipeline

  • Failed transmissions receive reconstructed coded-bit labels from later successful HARQ retransmissions, and retained samples use 16-QAM with MCS indices 6–10.For DUIDD finetuning and its Fig. 7 comparison, data are restricted to MCS index 10 and one scrambling configuration.Source: A. Finetuning Data Extraction Pipeline · pp. 3–4

B. Dataset Splits for Finetuning and Testing

  • November datasets use cross-UE finetuning and test splits, while June datasets are divided at a selected timestamp into temporally disjoint sets.Each set is constructed so about 10% of transport blocks fail with the MMSE Reference Rx; receivers in a comparison share test samples.Source: B. Dataset Splits for Finetuning and… · p. 4

C. Site-Specific Finetuning Pipeline

  • NRX and MDX are pretrained on synthetic 3GPP UMi channels, while DUIDD is pretrained on synthetic dual-layer channels with BCE followed by normalized LogSumExp BLER loss.The MDX pretraining uses single-user, single-layer transmission; DUIDD pretraining spans 6 · 10^3 batches.Source: C. Site-Specific Finetuning Pipeline · p. 4
  • NRX and MDX are finetuned with BCE on 10^5 batches of 16 slot samples cropped to four PRBs, while DUIDD uses 2 · 10^3 batches of 16 full-band samples.DUIDD uses BCE for the first 10^3 batches and normalized LogSumExp BLER loss for the next 10^3.Source: C. Site-Specific Finetuning Pipeline · p. 4

D. Site-Specific Performance Evaluation

The evaluation compares receiver performance on fixed measured transmissions and across effective SNR values. Site-specific finetuning lowers MDX dataset BLER in both single-layer environments, though the MMSE Reference Rx remains better.

D. Site-Specific Performance Evaluation

  • Dataset BLER is the fraction of transport blocks failing after offline receiver processing and LDPC decoding on a fixed test set, enabling fair comparisons on the same recordings.The test set is sampled so the MMSE Reference Rx achieves about 10% dataset BLER.Source: D. Site-Specific Performance Evaluation · p. 4
  • Adding complex Gaussian noise to the test data sweeps effective SNR downward and produces dataset-BLER curves.Effective SNR is computed from the original signal and receiver-noise powers plus the added-noise variance.Source: D. Site-Specific Performance Evaluation · p. 4

V. RECEIVER FINETUNING RESULTS

  • The NRX, MDX, non-iterative LMMSE, IDD, and DUIDD use standard PUSCH LS channel estimates with adjacent-pilot OCC combining; the MMSE Reference Rx uses multi-stage MMSE estimation with delay estimation.Source: §V · p. 4

1) Model-Driven Neural Receiver Finetuning:

  • Finetuning cuts MDX dataset BLER from 0.351 to 0.237 in the small lab and from 0.463 to 0.240 on the office floor; MMSE Reference Rx remains lower.The comparison uses single-layer measurements from June 2026.Source: 1) Model-Driven Neural Receiver… · p. 4

2) Temporal Robustness of NRX Finetuning:

NRX finetuning remains beneficial when training and testing use measurement campaigns more than half a year apart, although the temporal gap increases BLER. For dual-layer evaluation, finetuning on both single- and dual-layer data yields results close to dual-layer-only finetuning.

2) Temporal Robustness of NRX Finetuning:

  • Across November 2025 and June 2026, same-campaign finetuning gives the lowest dataset BLER, but either campaign’s finetuned NRX beats the pretrained model on the other campaign.The temporal gap increases BLER most visibly for the Shallow NRX tested on June 2026 data.Source: 2) Temporal Robustness of NRX Finetuning · pp. 4–5, Figure 4 · p. 5

1) Finetuning on Dual-Layer vs. Single- and Dual-Layer Data:

  • On dual-layer transmission, finetuning only on dual-layer data reduced dataset BLER by more than one third for both Shallow (2-iteration) and Deep (8-iteration) NRXs; adding single-layer data changed it by at most 0.019.Source: 1) Finetuning on Dual-Layer vs.… · p. 5, Figure 5 · p. 5
  • In a further experiment not presented in the paper, joint single- and dual-layer finetuning yields single-layer dataset BLER within 0.004 of single-layer-only finetuning.Source: 1) Finetuning on Dual-Layer vs.… · p. 5

2) Effective SNR Improvement of NRX Finetuning:

Site-specific NRX finetuning improves dual-layer effective-SNR performance, while DUIDD finetuning yields only a marginal additional BLER reduction. NRXs achieve lower BLER and larger finetuning gains than DUIDD, at higher complexity; the parameter-adaptation section introduces LMMSE channel estimation with covariance matrices estimated from synthetic or site-specific data.

2) Effective SNR Improvement of NRX Finetuning:

  • 0.83 dB for Shallow NRX and 0.54 dB for Deep NRX reduce the effective SNR required for 25% dataset BLER after dual-layer finetuning.The improvement holds throughout the evaluated waterfall region.Source: 2) Effective SNR Improvement of NRX… · p. 5

3) Model-Based DUIDD Receiver Finetuning:

  • DUIDD finetuning reduces dataset BLER by only 0.004 in absolute terms, whereas both NRX configurations achieve lower BLER and substantially larger finetuning gains at much higher complexity.DUIDD has similar complexity to classical IDD and about twice the detection complexity of the non-iterative baseline [4, Sec. V-D].Source: 3) Model-Based DUIDD Receiver Finetuning · p. 6, Figure 7 · p. 6

VI. SITE-SPECIFIC PARAMETER ADAPTATION

  • The LMMSE channel-estimation study compares covariance matrices estimated from synthetic channels with those estimated from site-specific measurements.This section investigates their impact on dataset BLER.Source: §VI · p. 6

A. LMMSE Channel Estimation

The receiver smooths direct LS channel estimates with a successive one-dimensional LMMSE interpolator. The method applies frequency-domain estimation and analogous spatial and temporal smoothing using channel covariance matrices.

A. LMMSE Channel Estimation

  • Sionna’s LMMSE interpolator approximates joint three-dimensional estimation through successive smoothing across receive antennas, subcarriers, then OFDM symbols.The order is explicitly antennas, frequency, and time.Source: A. LMMSE Channel Estimation · p. 6
  • The modified PUSCH LS estimator omits orthogonal cover-code averaging and produces direct estimates by dividing each received DMRS observation by its transmitted symbol.The estimate has channel coefficient h[w] and estimation error e[w].Source: A. LMMSE Channel Estimation · p. 6
  • Frequency-domain LMMSE estimation uses pilot-subcarrier observations selected from the full-band channel, with channel and error covariance matrices.The channel vector spans active subcarriers; spatial and temporal smoothing use their corresponding covariance matrices, while error covariances are derived at runtime from a noise-variance estimate.Source: A. LMMSE Channel Estimation · p. 6

B. Site-Specific Covariance Matrix Estimation

The study estimates covariance matrices from synthetic channels as a baseline and from site-specific measurements using noise-corrected second moments. Spatial estimation subtracts an estimated error covariance, while frequency estimation imposes a Toeplitz structure to infer correlations unavailable from pilot placement.

B. Site-Specific Covariance Matrix Estimation

  • The baseline covariance matrices are estimated from synthetic 3GPP UMi channel realizations by averaging channel outer products across randomized topologies and realizations.The implementation samples full-band OFDM-domain channels.Source: B. Site-Specific Covariance Matrix… · p. 6
  • Site-specific covariance estimation forms second moments from direct LS estimates and removes their estimation-error contribution, estimated using noise-only observations.The procedures use the finetuning data’s resource-grid configuration.Source: B. Site-Specific Covariance Matrix… · p. 6

1) Spatial Covariance Matrix:

  • Spatial covariance is estimated by averaging outer products of four-antenna LS estimate vectors, then subtracting the estimated spatial error covariance.Vectors are formed for each slot sample, layer, DMRS symbol, and pilot subcarrier.Source: 1) Spatial Covariance Matrix · pp. 6–7

2) Frequency Covariance Matrix:

  • Because pilots occupy every second subcarrier, the method assumes common marginal frequency statistics and imposes a Hermitian Toeplitz covariance structure.It estimates available even-lag correlations after error correction, interpolates missing odd lags from neighboring even lags, and uses Hermitian symmetry for negative lags.Source: 2) Frequency Covariance Matrix · p. 7

3) Temporal Covariance Matrix:

Temporal covariance estimation subtracts estimated LS-error covariance from channel-estimate second moments, while runtime limits the LMMSE evaluation to smaller test datasets.

3) Temporal Covariance Matrix:

  • The temporal covariance estimate subtracts estimated LS-error covariance from second moments of interpolated channel estimates.Channel estimates across all 13 OFDM symbols are stacked and averaged; negative covariance eigenvalues are set to zero to ensure positive semidefiniteness.Source: 3) Temporal Covariance Matrix · p. 7

C. Receiver Adaptation Results

  • Because Sionna's LMMSE estimator is slow, evaluations use smaller test datasets; absolute BLER comparisons need care, particularly across receivers using different LDPC decoders.Source: C. Receiver Adaptation Results · p. 7
  • Fig. 8 evaluates dataset BLER for a non-iterative LMMSE receiver across LS and LMMSE channel estimators on single- and dual-layer June 2026 datasets.The single-layer dataset is shown on the left and the dual-layer dataset on the right.Source: Figure 8 · p. 7, C. Receiver Adaptation Results · p. 7

1) Results with Non-Iterative Receivers:

Site-specific covariance estimation combined with iterative detection achieves the lowest dataset BLER reported in the study, while site-specific finetuning benefits neural receivers more than the less tunable model-based receiver.

1) Results with Non-Iterative Receivers:

2) Results with Iterative Receivers:

VII. CONCLUSIONS

  • Site-specific finetuning substantially benefits the fully trainable and model-driven neural receivers but yields only marginal gains for the less-tunable DUIDD receiver.A single NRX jointly finetuned for single- and dual-layer transmission has only minor performance losses relative to separately finetuned NRXs; finetuning remains effective across campaigns separated by more than half a year.Source: §VII · p. 8
Loading 2609.04004v1…