⚡ electrical engineering

Deep Learning assisted Port-Cycling based Channel Sounding for Precoder Estimation in Massive MIMO Arrays

This paper proposes a deep learning-based framework called CsiAdaNet that utilizes a port-cycling mechanism to reduce reference signal overhead while accurately reconstructing full-port channel state information for massive MIMO systems in future 6G networks.

Original authors: Advaith Arun, Shiv Shankar, Dhivagar Baskaran, Klutto Milleth, Bhaskar Ramamurthi

Published 2026-01-26

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Advaith Arun, Shiv Shankar, Dhivagar Baskaran, Klutto Milleth, Bhaskar Ramamurthi

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Big Problem: Too Many Microphones, Not Enough Time

Imagine a massive concert hall (the cell tower) trying to talk to a single listener (your phone). To make the sound perfect, the hall has 128 microphones (antennas) instead of just one. To figure out exactly how to aim the sound so it hits the listener clearly, the hall needs to "test" every single microphone.

In current technology, the hall has to shout a test signal through all 128 microphones at the exact same time. This takes up a huge amount of "airtime" (resources). It's like trying to interview 128 people simultaneously; you spend so much time just asking questions that you have no time left to actually have a conversation (send data). This slows down your internet speed.

The Proposed Solution: The "Port-Cycling" Strategy

The authors propose a smarter way to do this test, which they call Port-Cycling.

Instead of shouting through all 128 microphones at once, the system activates them in small groups.

The Analogy: Imagine a security guard checking a large warehouse. Instead of trying to see the whole room at once (which is impossible), the guard shines a flashlight on the left corner, then the right corner, then the back, one by one.
How it works: The system turns on a small group of antennas, listens for a moment, turns them off, and turns on the next group. It cycles through all the groups over a short period.
The Benefit: At any single instant, the system is only using a fraction of the microphones. This saves a massive amount of "airtime" (overhead), leaving more room for actual data to be sent.

The Catch: Can We Reconstruct the Whole Picture?

If you only look at the corners of the room one by one, how do you know what's happening in the middle? You might miss the big picture.

This is where Deep Learning (AI) comes in. The authors built a smart computer model called CsiAdaNet.

The Analogy: Think of a detective who only sees a few blurry snapshots of a crime scene taken at different times. A normal person might be confused. But this "AI Detective" has studied thousands of crime scenes before. It knows that if the shadow in the left corner looks a certain way, and the sound in the right corner sounds a certain way, the object in the middle must be there.
How it works: The AI looks at the small, partial measurements taken over time. It uses its "training" to understand how the signal moves through space and time. It fills in the missing gaps to reconstruct the full picture of the channel, just as if all 128 microphones had been tested at once.

How the AI Works (The "CsiAdaNet" Model)

The model doesn't just guess; it follows a specific hierarchy, like solving a puzzle step-by-step:

Find the General Area: First, it guesses the general "beamset" (the broad direction of the signal).
Pinpoint the Target: Next, it picks the specific "beam indices" (the exact antennas to focus on).
Adjust the Volume and Timing: Finally, it calculates the precise "amplitude" (volume) and "phase" (timing) needed to combine the signals perfectly.

The model is trained on a massive dataset of simulated signals so it learns the patterns of how radio waves behave in a city environment.

The Results: Fast and Accurate

The authors tested this system in a simulated city environment (Urban Macro) with a 128-antenna tower.

Performance: The AI's reconstruction was almost identical to the "perfect" method where all antennas are tested at once.
Accuracy: In good conditions (moderate to high signal strength), the AI predicted the correct settings with very high accuracy.
Efficiency: The system successfully reduced the immediate "overhead" (the time spent testing) by a factor of 4, without losing much performance.
Complexity: The AI model is actually computationally lighter than the traditional math-heavy methods used today, making it easier to run on future hardware.

Summary

In short, this paper proposes a way to talk to massive antenna arrays without wasting time. Instead of testing every antenna at once, the system tests them in a rotating cycle. A smart AI then takes these partial, rotating snapshots and stitches them together to create a perfect, full-resolution map of the connection, allowing for faster and more efficient 6G-style wireless communication.

Technical Summary: Deep Learning Assisted Port-Cycling based Channel Sounding for Precoder Estimation in Massive MIMO Arrays

Problem Statement
Future 6G wireless systems are expected to utilize a significantly larger number of transmit ports for Channel State Information (CSI) estimation compared to current specifications. While this increases spectral efficiency, it creates a substantial resource overhead for transmitting reference signals (CSI-RS) across the time-frequency grid, ultimately reducing achievable data throughput. Existing AI/ML approaches for CSI feedback, such as CsiNet, primarily address feedback compression or temporal extrapolation but fail to mitigate the instantaneous overhead caused by large-scale antenna sounding. Furthermore, traditional environment-based prediction methods relying on site-specific information (e.g., images) lack generalizability. The core challenge addressed in this work is reducing the instantaneous spatial overhead in Massive MIMO systems while maintaining compatibility with current standard-compliant transmission and preserving channel observability.

Methodology
The proposed solution integrates a novel channel sounding strategy with a deep learning (DL) framework:

Port-Cycling Mechanism: Instead of activating all antenna ports simultaneously, the Uniform Planar Array (UPA) is partitioned into non-overlapping sub-panels. These sub-panels are sequentially sounded across multiple time instances (cycles). This reduces the instantaneous number of active ports by a factor of $\rho$ (the spatial reduction factor) while ensuring the full array is covered over time. The feasibility of this approach relies on the temporal coherence of beam-related quantities (beamset and dominant beam index) during the cycling duration, which the authors validate via a variation score analysis.
CsiAdaNet Model: A deep learning model, termed CsiAdaNet, is deployed at the User Equipment (UE) to reconstruct full-port CSI from the temporally aggregated, sparse port-cycled measurements. The model operates hierarchically to predict the quantized Type-II precoder parameters defined in 5G NR:
- Input: A sequence of $\rho$ CSI tensors corresponding to the sub-arrays sounded at different time steps.
- Architecture: The model employs a shared encoder module using 3D convolutional layers to extract spatial-frequency features from each sub-array. A Gated Recurrent Unit (GRU) processes the sequence of encoded features to capture temporal correlations.
- Output Heads: The aggregated representation is fed into four task-specific heads:
  - Beam Set Head: Predicts the over-sampled beamset index.
  - Beam Indices Head: Predicts the top- $L$ orthogonal beam indices within the selected set.
  - Wideband Amplitude Head: Predicts quantized amplitude levels.
  - Wideband Phase Head: Predicts quantized phase levels.
- Training: The model is trained using a weighted multi-task sparse categorical cross-entropy loss, with higher weights assigned to beam indices to minimize error propagation in precoder construction.

Key Contributions
The paper outlines three primary contributions:

Port-Cycling Sounding Technique: A method to reduce instantaneous spatial overhead in Massive MIMO by sequentially activating sub-panels, preserving full array observability over time.
CsiAdaNet Architecture: A deep learning model capable of directly predicting Type-II precoder quantities (beamset, indices, amplitude, and phase) from temporally aggregated port-cycled measurements, eliminating the need for full-port instantaneous CSI.
Link-Level Validation: Demonstration of the approach's performance in a 3GPP-compliant simulation setup, showing near-optimal performance compared to conventional methods.

Simulation Results
The authors evaluated the system using a 3GPP-compliant link-level simulator under an Urban Macro (UMa) Line-of-Sight scenario with a 128-antenna BS and 4-antenna UE. Key findings include:

Precoder Alignment: The CsiAdaNet approach closely matches the conventional baseline (full-port sounding) across all Signal-to-Noise Ratio (SNR) levels. The Squared Generalized Cosine Similarity (SGCS) increases from below 0.2 at low SNR to approximately 0.7 at medium-to-high SNR.
Beamforming Gain: The gain achieved using CsiAdaNet exhibits a close match to the traditional baseline, with a negligible gap (within $\approx 0.03$ ) at high SNR.
Component Accuracy:
- Beamset Prediction: Highly reliable even at moderate SNR due to coarse spatial resolution.
- Beam Indices: Accuracy improves gradually with SNR, reflecting finer angular sensitivity.
- Amplitude and Phase: Both show high accuracy at moderate-to-high SNR, though phase prediction is more challenging at low SNR.
Complexity: Unlike conventional methods requiring Eigenvalue Decomposition (EVD) of the full covariance matrix with $O(N_t^3)$ complexity, CsiAdaNet performs inference on sub-panel observations with complexity $O(\rho D^2)$ , offering improved scalability for large arrays.

Significance and Claims
The paper claims that this work provides a viable enabler for reliable CSI acquisition in future 6G systems. By reducing the number of active ports at any given instant by a factor of $\rho$ , the method significantly lowers instantaneous overhead and RF chain usage without sacrificing long-term channel observability. The authors assert that the proposed DL-assisted framework achieves near-optimal beam alignment and beamforming gain at moderate to high SNR levels, with only marginal degradation in noisy regions. This approach highlights a key design consideration for scalable and energy-efficient CSI acquisition in next-generation large-scale MIMO systems, bridging the gap between the high overhead of massive antenna arrays and the practical constraints of reference signal transmission.