Abstract
Load-modulated Multiple-Input Multiple-Output (MIMO) systems offer significant energy efficiency advantages by utilizing a single radio frequency (RF) chain to drive multiple antennas. However, signal detection for Phase Modulation on the Hypersphere (PMH) signals in these systems is hindered by the need for accurate Channel State Information (CSI) and the exponential computational complexity of Maximum Likelihood (ML) detection. This paper presents two machine learning-based detection schemes: Hypersphere Clustering via Expectation-Maximization (HEM-ML) and EM with KD-tree Detection (HEM-KD). By framing signal detection as an unsupervised clustering problem, the HEM-ML scheme jointly estimates CSI and detects signals with minimal pilot overhead. The HEM-KD scheme further optimizes the process by employing KD-tree structures to transform the detection into a linear-time nearest neighbor search. Simulation results show that both schemes achieve near-optimal performance while drastically reducing computational overhead compared to traditional methods.
Introduction
The demand for high-capacity wireless communication has made MIMO technology indispensable for 5G and beyond. Traditional MIMO architectures, however, suffer from high power consumption due to the requirement of dedicated RF chains and power amplifiers for each antenna. Load-modulated MIMO addresses this by using a single central power amplifier and RF chain, modulating signals through variable loads at each antenna. While efficient, this architecture complicates signal detection, particularly when using Phase Modulation on the Hypersphere (PMH). Conventional ML detection requires precise CSI and faces a "complexity explosion" as the number of antennas and bits per symbol increase.
System Architecture
The load-modulated MIMO system utilizes a single RF chain to feed multiple antennas, where each antenna is equipped with a load modulator. This configuration significantly reduces hardware complexity and power consumption compared to conventional MIMO systems.
Proposed Machine Learning Schemes
To address the challenges of CSI acquisition and detection complexity, the authors propose a two-stage machine learning approach:
Hypersphere Clustering via EM (HEM-ML)
This scheme treats signal detection as an unsupervised clustering problem. Since PMH signals naturally form clusters on a multidimensional hypersphere at the receiver, they can be modeled using a Gaussian Mixture Model (GMM). The Expectation-Maximization (EM) algorithm is employed to "learn" the channel features. By using a small number of pilot symbols for initialization, the algorithm jointly refines the CSI and signal decisions using received data symbols, thereby minimizing pilot overhead.
EM with KD-tree Detection (HEM-KD)
To solve the exponential complexity of brute-force ML detection, the HEM-KD scheme replaces the search process with a KD-tree data structure. The CSI learned during the EM stage is used to build a spatial KD-tree codebook. Signal detection is then transformed into a Nearest Neighbor Search (NNS). By traversing the tree, the receiver can identify the correct symbol in linear time, providing a massive computational speedup without sacrificing accuracy.
Performance Evaluation
The proposed methods were evaluated using Rayleigh fading and Millimeter Wave (mmWave) channel models. Key findings include:
- Near-Optimal Accuracy: Both HEM-ML and HEM-KD achieve Bit Error Rate (BER) performance nearly identical to the "perfect CSI" scenario as the Signal-to-Noise Ratio (SNR) increases.
- Computational Efficiency: In a test case with $N=3$ antennas, the HEM-KD scheme required only 19,877 operations per frame, compared to over 27 million for a high-overhead Least Squares (LS) approach.
- Reduced Overhead: The schemes significantly outperform traditional LS estimation when using the same limited number of pilot symbols, demonstrating superior resource efficiency.
Conclusion
The integration of machine learning techniques—specifically EM-based clustering and KD-tree search—provides a robust solution for signal detection in load-modulated MIMO systems. These methods effectively bypass the CSI bottleneck and complexity challenges, offering a path toward more energy-efficient and scalable next-generation wireless transceivers. Future work may involve extending these models to highly mobile, time-varying channel environments.
Summary updated