⚡ electrical engineering

ChannelKAN: Multi-Scale Dual-Domain Channel Prediction via Hybrid CNN-KAN Architecture

The paper proposes ChannelKAN, a hybrid CNN-KAN architecture with multi-scale dual-domain enhancement that outperforms existing deep learning baselines in predicting channel state information for high-mobility massive MIMO-OFDM systems by effectively capturing both local spatial-frequency correlations and long-range nonlinear temporal dependencies.

Original authors: Nanqing Jiang, Zhangyao Song, Tao Guo, Xiaoyu Zhao, Yinfei Xu

Published 2026-05-14

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Nanqing Jiang, Zhangyao Song, Tao Guo, Xiaoyu Zhao, Yinfei Xu

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

Imagine you are trying to predict the weather for the next few hours. If you just look at the sky right now, you might miss a storm brewing far away. If you only look at historical weather patterns from last year, you might miss a sudden, local gust of wind. To get it right, you need to understand both the immediate, local changes and the big, long-term trends.

This is exactly the problem the paper "CHANNELKAN" is trying to solve, but instead of weather, they are predicting wireless signals (specifically for 5G and future 6G networks) in a world where everything is moving fast—like cars on a highway or trains.

Here is a simple breakdown of what they did and why it works:

The Problem: The "Aging" Signal

When you send data over the air, the signal bounces off buildings, trees, and cars. This creates a "map" of the connection called Channel State Information (CSI).

The Issue: In a fast-moving world, this map changes incredibly quickly. By the time the phone calculates the map and tries to use it, the map is already "old" (outdated). This is called channel aging.
The Goal: The researchers want to build a crystal ball that can predict what the signal map will look like a few moments in the future, so the network can stay fast and reliable even when you are driving at 100 km/h.

The Old Ways vs. The New Way

Previous methods tried to solve this with two main tools, but both had flaws:

RNNs/LSTMs (The "Step-by-Step" Walkers): These look at the signal one moment at a time. They are good at seeing what happens right now, but they get confused and forget things when trying to remember long-term patterns. It's like trying to read a book by only looking at one letter at a time; you miss the story.
Transformers (The "Global" Scanners): These look at the whole picture at once. They are great at seeing long patterns but struggle to catch the tiny, rapid, local shifts that happen in a split second. They are like looking at a satellite photo of a forest; you see the whole forest, but you miss the single leaf falling.

The Paper's Solution: CHANNELKAN
The authors built a hybrid system that combines the best of both worlds, using two special types of AI: CNNs and KANs.

1. The "Dual-Domain" Expansion (Looking at the Signal Twice)

Imagine you are listening to a song. You can listen to the melody (the notes over time) or the instrumentation (the specific frequencies).

The researchers take the signal and create two versions of it: one in the Frequency Domain (like the melody) and one in the Delay Domain (like the echoes or how long it takes to bounce back).
By feeding both versions into the AI, they give the model two different "lenses" to see the same reality, making it much harder to miss important details.

2. The "Multi-Scale" Filter (Cleaning the Noise)

Wireless signals are often noisy, like trying to hear a conversation in a loud bar.

The model uses a special filter (based on math called FFT) to look at the signal at different "scales."
It acts like a smart noise-canceling headphone: it keeps the loud, important parts of the signal (the dominant frequencies) and throws away the tiny, useless static and noise. This ensures the AI isn't distracted by garbage data.

3. The Hybrid Brain: CNN + KAN

This is the core innovation. The model splits the work between two specialized "brains":

The CNN (The Local Detective): This part uses Convolutional Neural Networks. Think of it as a detective with a magnifying glass. It zooms in on tiny, local details within a single moment to see how the signal is shifting right now. It's excellent at spotting immediate, short-term changes.
The KAN (The Long-Range Storyteller): This part uses Kolmogorov-Arnold Networks. Unlike standard AI that uses fixed rules, KANs use flexible, learnable mathematical curves (Chebyshev polynomials). Think of this as a storyteller who looks at the entire sequence of events and understands the complex, non-linear story of how the signal evolves over time. It connects the dots across the whole timeline.

The Magic: The CNN handles the "what's happening right now," and the KAN handles "where is this going." They work together to predict the future signal perfectly.

4. The Fusion (Putting it Together)

Finally, the model takes the insights from the "Local Detective" and the "Long-Range Storyteller," combines them, and produces a prediction of what the signal will look like in the future.

The Results: Why It Matters

The researchers tested this on a massive, realistic simulator that mimics 3GPP (the global standard for mobile networks) conditions.

Speed: They tested speeds from 10 km/h (walking) to 100 km/h (high-speed train).
Noise: They tested in both clear conditions and very noisy environments.

The Outcome:
CHANNELKAN beat every other method they compared it to (including RNNs, LSTMs, GRUs, CNNs, and Transformers).

It made fewer mistakes (lower error rate).
It allowed for more data to be sent (higher speed/efficiency).
It made fewer connection drops (lower error rate in data transmission).

In a Nutshell:
The paper claims that by combining a "magnifying glass" (CNN) for local details with a "storyteller" (KAN) for long-term trends, and by cleaning up the signal first, they created a system that can predict wireless signals much more accurately than anything else, especially when things are moving fast. This means your phone could stay connected and fast even on a speeding train.

Technical Summary: CHANNELKAN

Problem Statement

Accurate Channel State Information (CSI) prediction is critical for enhancing the reliability and spectral efficiency of massive MIMO-OFDM systems, particularly in high-mobility scenarios. In Time-Division Duplex (TDD) systems, channel aging caused by rapid fluctuations (Doppler shifts) and multipath propagation renders estimated CSI outdated before it can be utilized for resource allocation.

Existing deep learning approaches face a fundamental limitation: they struggle to jointly capture two distinct types of temporal dependencies required for accurate prediction:

Short-term local variations: Rapid fluctuations between adjacent time steps.
Long-range nonlinear dependencies: The overall temporal evolution across the entire observation window.

Specific architectures have inherent drawbacks:

RNNs (LSTM/GRU): Effective for short-term correlations but suffer from gradient vanishing when modeling long-range dependencies.
CNNs: Efficient at extracting local spatial-frequency features but are restricted by fixed receptive fields, limiting their ability to capture dependencies beyond local neighborhoods.
Transformers: Capable of modeling global dependencies via self-attention but suffer from quadratic complexity and rely on linear projections that may not adequately fit the highly nonlinear temporal evolution of wireless channels.

Methodology

The authors propose CHANNELKAN, a hybrid architecture that integrates Convolutional Neural Networks (CNNs) and Kolmogorov-Arnold Networks (KANs) to decouple and jointly model short-term and long-range dependencies. The model operates on 3GPP-compliant channel datasets generated by the QuaDRiGa simulator and consists of four primary modules:

1. Dual-Domain Expansion Module (DDE)

This module generates complementary representations of the CSI to provide richer input features:

Frequency Domain Representation (CFR): The input is flattened to preserve subcarrier-domain correlations.
Delay Domain Representation (CIR): The CSI is transformed via Inverse Discrete Fourier Transform (IDFT) to reveal sparse multipath components.
The features from both domains are concatenated to form a dual-domain feature vector.

2. Multi-Scale Frequency Information Enhancement Module (MSFE)

To suppress noise and redundant fluctuations while preserving key patterns, this module:

Applies a real-valued Fast Fourier Transform (rFFT) along the temporal dimension.
Selects dominant frequency components at multiple scales using a Top-K selection strategy.
Constructs binary masks to filter out trivial components and reconstructs the enhanced time-domain features via inverse FFT (irFFT).
Aggregates multi-scale features using learnable dense layers.

3. Dual-Domain CNN-KAN Feature Extraction Module (DFE)

This core component processes the enhanced features through parallel branches for both CFR and CIR domains:

Local Feature Extraction (CNN): Cascaded convolutional layers extract intra-time-step local spatial-frequency correlations.
Long-Range Dependency Modeling (KAN): The output is reshaped into a 2D matrix and processed by Chebyshev KAN layers. Unlike standard fully connected layers with fixed activations, KANs place learnable univariate functions (specifically Chebyshev polynomials) on network edges. This allows for holistic nonlinear mapping of the entire feature matrix to model inter-time-step temporal evolution.

4. Dual-Domain Fusion Module (DDF)

Features from the CFR and CIR branches are concatenated and passed through a dense fusion layer to produce the final prediction. The output is reshaped and inversely transformed to recover the complex-valued predicted CSI.

Key Contributions

The paper outlines four main contributions:

Hybrid CNN-KAN Architecture: The first application of KANs to channel prediction. It explicitly decouples modeling tasks: CNNs handle short-term local spatial-frequency correlations, while Chebyshev polynomial-based KANs model long-range nonlinear temporal evolution.
Multi-Scale Frequency Enhancement: A module that uses FFT-based filtering to select dominant spectral components at multiple scales, effectively strengthening key channel features and suppressing noise.
Dual-Domain Framework: A processing and fusion framework that jointly exploits frequency-domain and delay-domain CSI representations to provide complementary information and enhance robustness.
Comprehensive Validation: Extensive experiments on 3GPP-compliant datasets validating the model across diverse velocities and Signal-to-Noise Ratios (SNRs), supported by ablation studies confirming the necessity of each module.

Experimental Results

Experiments were conducted on QuaDRiGa datasets with a historical window of $T=16$ and a prediction horizon of $L=4$ . The model was compared against RNN, LSTM, GRU, CNN, and Transformer baselines.

NMSE Performance: CHANNELKAN consistently achieved the lowest Normalized Mean Square Error (NMSE) across all tested velocities (10–100 km/h). It showed a significant advantage at high velocities (e.g., NMSE $\approx$ 0.0609 at 100 km/h), where baseline models (particularly RNNs and Transformers) degraded rapidly.
Robustness to Noise: The model outperformed all baselines across SNR levels (0–30 dB), with notable superiority at low SNR levels, attributed to the multi-scale frequency enhancement module.
System Metrics: CHANNELKAN achieved the highest Spectral Efficiency (SE) of 6.414 bps/Hz and the lowest Bit Error Rate (BER) of 0.00771.
Efficiency: Despite superior performance, the model maintained a moderate parameter size of 24.5 MB, which is competitive compared to the Transformer baseline (65 MB).

Significance and Claims

The paper claims that CHANNELKAN successfully addresses the "joint capture" challenge that existing deep learning methods fail to solve. By leveraging the complementary nature of CNNs (local extraction) and KANs (holistic nonlinear fitting), the model achieves a balance between efficiency and flexibility.

The authors emphasize that this is the first application of KANs to channel prediction. The results suggest that decoupling short-term and long-term dependency modeling leads to superior performance in high-mobility environments. The paper concludes that while the model is effective, future work is required to reduce computational costs for real-time deployment and to validate generalization on measured (real-world) channel data.