Source-linked AI summary
Hardware-in-the-Loop Phase-Aware CNN for Real-Time 5G Channel Estimation
Javad Zolfaghari-Bengar, Rakibul Rony, Elisa Gomez-de-Lope, Alejandro Villena-Rodriguez, Abhinav Mahadevan, Nicolas Kourtellis
TL;DR
Hardware impairments complicate accurate 5G channel estimation, motivating evidence beyond simulation-only evaluations. This demo uses a hardware-in-the-loop phase-aware CNN for real-time DMRS-based reconstruction, reducing magnitude MSE against LS from 0.0860 to 0.0105.
Problem
Accurate 5G channel estimation must account for propagation, noise, RF impairments, synchronization, radio-unit processing, and baseband implementation details.
Method
A lightweight phase-aware CNN estimates channel magnitude and phase from hardware-derived uplink DMRS observations and is compared with LS and oracle frequency-domain LMMSE baselines.
Results
Magnitude MSE falls from 0.0860 with LS to 0.0105 with the CNN, while the CNN achieves the lowest MSE across the evaluated SNR range.
Takeaways & Limitations
The demo provides a practical real-time example of AI-based channel estimation using realistic hardware-derived 5G uplink data.
Abstract
from arXiv · showhide
This demo presents real-time AI-based uplink channel-estimation inference using data collected from a hardware-in-the-loop 5G platform. The data-collection setup integrates commercial RF signal generation, programmable channel emulation, an O-RAN Radio Unit, DU emulation, and a lightweight phase-aware convolutional neural network (CNN) that estimates the channel response directly from received DMRS signals. Unlike simulation-only evaluations, the hardware-derived dataset exposes the estimator to practical RF and system-level impairments, including calibration mismatches, synchronization imperfections, quantization effects, phase noise, and implementation-specific nonlinearities. During the demo, attendees will observe real-time CNN inference and channel reconstruction using captured hardware-generated DMRS observations and compare the proposed CNN against Least Squares (LS) and frequency-domain LMMSE baselines. The objective is to showcase a practical AI-native physical-layer inference pipeline that combines hardware-derived 5G data with real-time neural channel estimation for future 5G-Advanced and 6G systems.
I. INTRODUCTION
The section motivates hardware-aware channel estimation for reliable 5G/6G physical-layer processing and presents a real-time AI inference pipeline based on hardware-in-the-loop DMRS data. The demo combines commercial RF, programmable channel emulation, O-RAN and DU processing, and compares a lightweight phase-aware CNN with LS and frequency-domain LMMSE estimators.
- Motivation: Accurate channel estimation supports reliable demodulation, interference management, and beamforming in 5G and emerging 6G systems.The need grows as deployments become denser and channels more dynamic.
- Motivation: Received pilots are affected by propagation, noise, RF impairments, synchronization behavior, radio-unit processing, and baseband implementation details.These effects motivate estimators designed for practical system conditions.
- Estimator context: LS is simple and fast but propagates noise, whereas MMSE/LMMSE requires covariance information that may be unavailable or difficult to maintain in dynamic hardware-impaired environments.The introduction identifies deep-learning estimators as promising for OFDM and MIMO channel estimation.
- Platform and workflow: The demo uses received DMRS signals from a hardware-in-the-loop platform combining commercial RF generation, programmable channel emulation, O-RAN radio processing, and DU emulation.The workflow is presented as an AI-native physical-layer pipeline for future O-RAN and 6G systems.
- Demo evaluation: Captured hardware-derived data are processed by a lightweight phase-aware CNN in real time and compared with LS and frequency-domain LMMSE estimators.The comparison occurs during the conference demo.
II. DEMONSTRATION PLATFORM
The demonstration platform collects uplink DMRS data through a hardware-in-the-loop 5G chain spanning RF generation, channel emulation, O-RAN radio conversion, and DU processing. This hardware-derived setup exposes the learning model to practical RF effects that pure software simulation may not capture.
- Signal generation and channel emulation: Two independent single-antenna UEs are emulated with Keysight MXG N5182B RF generators producing uplink DMRS signals.The signals pass through controlled 5G propagation scenarios in a Keysight PROPSIM F8800B channel emulator.
- O-RAN and DU processing: A Liteon O-RAN Radio Unit performs RF-to-baseband conversion before a Keysight S5040A DU emulator processes the resulting baseband I/Q samples through the 5G uplink chain.The received signals include PUSCH and DMRS I/Q data.
- AI platform and practical impairments: The platform supports AI training and real-time inference while exposing the model to calibration offsets, quantization effects, synchronization imperfections, and phase noise.These hardware-derived effects are described as difficult to capture with pure software simulation.
III. AI PROCESSING PIPELINE
The pipeline normalizes captured received DMRS observations, encodes magnitude and phase with sine–cosine components, and processes them through a lightweight CNN for real-time channel estimation. This representation supports stable joint magnitude-and-phase regression while avoiding wrapped-phase discontinuities.
- Preprocessing and inference: The preprocessing flow normalizes observations by transmitted DMRS, reshapes the sine–cosine encoding, and feeds it to the CNN for real-time channel reconstruction.The complete flow is illustrated in Fig. 2.
- Input representation: Captured received DMRS values are represented using their magnitude and phase encoded as sine and cosine components.The CNN operates on these transformed complex observations.
- Phase handling: Sine–cosine phase encoding avoids the artificial discontinuity of wrapped phase at ±π and enables stable regression of magnitude and phase.The estimated phase is recovered during inference.
- CNN architecture: The lightweight CNN uses five 2D convolutional layers with batch normalization to learn correlations across receive antennas, UEs, and subcarriers.Its footprint is under 1 MB, supporting real-time deployment on edge hardware.
- Training and baselines: Training uses Adam with learning rate 0.001, batch size 32, and 50 epochs, while evaluation includes an LS baseline.The passage states that the CNN is compared against two baselines.
IV. DEMO SCENARIO AND EVALUATION SETUP
The hardware-in-the-loop platform generated and validated a hardware-derived dataset across nine representative 3GPP scenarios, enabling evaluation of channel-reconstruction accuracy, inference latency, and throughput. The demo provides real-time CNN reconstruction and visualization from captured DMRS observations, with comparison against LS and frequency-domain LMMSE baselines.
- The hardware-in-the-loop platform generated and validated the dataset across nine representative 3GPP scenarios while evaluating reconstruction accuracy, inference latency, and throughput.
- Captured hardware-derived DMRS observations drive real-time CNN channel reconstruction and visualization of estimated magnitude and phase, alongside LS and frequency-domain LMMSE comparisons.
- The hardware-derived traces cover RMa, UMa, and UMi environments under indoor, outdoor LOS, and outdoor NLOS conditions.
- Each scenario used 12,000 training samples, 1,000 validation samples, and 2,000 testing samples.
V. REPRESENTATIVE RESULTS
The phase-aware CNN achieves the lowest channel-estimation MSE across the evaluated SNR range on hardware-derived channel realizations, with its largest gains at low and moderate SNR. Against LS, it reduces magnitude MSE from 0.0860 to 0.0105 and phase-domain MSE from 1.9554 to 0.7822.
- SNR-sweep evaluation: The phase-aware CNN achieves the lowest MSE across the evaluated SNR range against LS and oracle frequency-domain LMMSE estimators.The evaluation uses hardware-derived channel realizations.
- SNR-sweep evaluation: The largest CNN gains occur at low and moderate SNR values, where noise has the strongest impact.
- Error metrics: 0.0860 to 0.0105: the CNN reduces magnitude MSE against the LS baseline, corresponding to nearly an order-of-magnitude improvement.
- Error metrics: 1.9554 to 0.7822: the CNN reduces phase-domain MSE, indicating smoother and more accurate reconstruction of magnitude and phase.
VI. CONCLUSION
The demo demonstrates real-time AI-based channel-estimation inference on hardware-derived 5G uplink data, combining commercial RF equipment, O-RAN components, and controlled channel emulation. Attendees can inspect CNN inference, compare it with classical baselines, and observe real-time channel reconstruction.
- Demonstration: The demo performs real-time AI-based channel-estimation inference using data collected from a hardware-in-the-loop 5G uplink platform.The dataset is hardware-derived rather than simulation-only.
- Demonstration: The hardware setup combines commercial RF equipment, O-RAN components, and controlled channel emulation to produce realistic channel-estimation data.This integration supports data collection under practical platform conditions.
- Demonstration: Attendees can inspect CNN inference and visualization, compare the CNN with classical baselines, and observe real-time channel reconstruction.The conference demo presents these capabilities using captured hardware-derived data.