Abstract
The rapid growth of data traffic and the emerging AI-native wireless architectures in NextG cellular systems place new demands on the fronthaul links of Cloud Radio Access Networks (C-RAN). In this paper, we investigate neural compression techniques for the Common Public Radio Interface (CPRI), aiming to reduce the fronthaul bandwidth while preserving signal quality. We introduce two deep learning-based compression algorithms designed to optimize the transformation of wireless signals into bit sequences for CPRI transmission. The first algorithm utilizes a non-linear transformation coupled with scalar/vector quantization based on a learned codebook. The second algorithm generates a latent vector transformed into a variable-length output bit sequence via arithmetic encoding, guided by the predicted probability distribution of each latent element. Novel techniques such as a shared weight model for storage-limited devices and a successive refinement model for managing multiple CPRI links with varying Quality of Service (QoS) are proposed. Extensive simulation results demonstrate notable Error Vector Magnitude (EVM) gains with improved rate-distortion performance for both algorithms compared to traditional methods. The proposed solutions are robust to variations in channel conditions, modulation formats, and noise levels, highlighting their potential for enabling efficient and scalable fronthaul in NextG AI-native networks as well as aligning with the current 3GPP research directions.
Introduction
In Cloud Radio Access Network (C-RAN) architectures, baseband processing is centralized in Baseband Units (BBUs), while Remote Radio Heads (RRHs) handle radio frequency functions. These components communicate via the fronthaul using the Common Public Radio Interface (CPRI). As 6G approaches, the massive volume of raw I/Q signal data creates a significant bottleneck. Traditional compression methods like Scalar Quantization (SQ) often assume signal samples are independent and identically distributed (i.i.d.). However, mathematical analysis reveals that decimated CPRI signals are highly correlated, a property traditional linear methods fail to exploit, leading to bandwidth inefficiency or degraded signal quality.
Proposed Methodology
To address these inefficiencies, this research proposes two "AI-native" algorithms leveraging bi-directional Gated Recurrent Units (bi-GRUs) to capture temporal dependencies in wireless signals.
Neural Compression Algorithms
- Neural Latent Quantization: This approach transforms input signals into a latent space via a neural network, followed by quantization using a learned codebook optimized during training.
- Neural Compression with Probability Prediction: This advanced method uses a probability prediction network to estimate the distribution of latent variables, enabling arithmetic encoding. This entropy-based coding pushes the bitrate toward the theoretical limit of the latent representation's entropy.
Neural Architecture
The core of these algorithms is the bi-GRU architecture, which is specifically designed to handle sequences with internal correlations, moving beyond the flawed i.i.d. assumption of traditional methods.
System-Level Innovations
- Successive Refinement: For scenarios with multiple fronthaul links of varying reliability, this model sends a "base" coarse signal over the most reliable link and "refinement" bits over others to incrementally improve quality.
- Variable-Rate Compression: To accommodate hardware-constrained RRHs, a shared-weight model was developed. By sharing neural weights across different bitrates and using small lookup tables, the storage footprint is significantly reduced.
Experimental Setup
Models were evaluated using simulated datasets based on 3GPP 5G standards, featuring a 20 MHz bandwidth, 60 kHz subcarrier spacing, and a 512-point FFT. The channel model included a noisy 7-tap multi-path fading channel, with testing across 4QAM, 16QAM, and 64QAM modulation formats. Training utilized the Adam optimizer on batches of SC-FDMA and OFDM symbols.
Key Results and Discussion
The proposed neural algorithms consistently outperformed traditional SQ and Vector Quantization (VQ) benchmarks. In downlink scenarios, the neural models achieved significantly lower Error Vector Magnitude (EVM) across all tested bitrates. Specifically, the probability-based model achieved a reduction of approximately 0.1 bit per element over standard VQ while maintaining higher reconstruction fidelity.
Furthermore, the models demonstrated remarkable robustness. A model trained at 5dB SNR maintained high performance when tested at extremes (-5dB or 15dB) and remained effective despite changes in the number of channel taps. The use of non-linear neural transformations allows for complex boundaries that better fit the distribution of wireless signals compared to traditional linear partitioning.
Conclusion and Future Work
While neural compression offers superior rate-distortion performance and storage efficiency through weight sharing, challenges remain. The computational complexity of bi-GRUs is higher than simple scalar quantizers, necessitating further research into efficient FPGA/ASIC implementations for real-time, ultra-high-speed fronthaul. Future work will also explore extending these techniques to multi-user MIMO environments and conducting detailed hardware latency analyses to meet the strict requirements of NextG networks.
Summary updated