Explainable Autoencoder Design for RSSI-Based Multi-User Beam Probing and Hybrid Precoding

TL;DR

The paper addresses the overhead and black-box nature of beam management and hybrid precoding in MU-mmWave systems. It proposes Auto-HP, an end-to-end autoencoder that jointly learns site-specific probing beams and RF/BB precoders from limited RSSI measurements in an unsupervised setting, using complex-valued layers and RF constraints. An information-theoretic explainability framework is introduced to select the bottleneck dimension by maximizing entropy and mutual information, enhancing transparency in the learned probing beams. On the DeepMIMO DeepMIMO 3D ray-tracing dataset, the approach with as few as $M_{ ext{BS}}=8$ probing beams outperforms traditional $64$-DFT and $128$-O-DFT codebooks, reducing beam training overhead by up to $75\%$ and $87.5\%$, respectively, while achieving strong MU rates.

Abstract

This paper introduces a novel neural network (NN) structure referred to as an ``Auto-hybrid precoder'' (Auto-HP) and an unsupervised deep learning (DL) approach that jointly designs \ac{mmWave} probing beams and hybrid precoding matrix design for mmWave multi-user communication system with minimal training pilots. Our learning-based model capitalizes on prior channel observations to achieve two primary goals: designing a limited set of probing beams and predicting off-grid \ac{RF} beamforming vectors. The Auto-HP framework optimizes the probing beams in an unsupervised manner, concentrating the sensing power on the most promising spatial directions based on the surrounding environment. This is achieved through an innovative neural network architecture that respects \ac{RF} chain constraints and models received signal strength power measurements using complex-valued convolutional layers. Then, the autoencoder is trained to directly produce RF beamforming vectors for hybrid architectures, unconstrained by a predefined codebook, based on few projected received signal strength indicators (RSSIs). Finally, once the RF beamforming vectors for the multi-users are predicted, the baseband (BB) digital precoders are designed accounting for the multi-user interference. The Auto-HP neural network is trained end-to-end (E2E) in an unsupervised learning manner with a customized loss function that aims to maximizes the received signal strength. The adequacy of the Auto-HP NN's bottleneck layer dimension is evaluated from an information theory perspective, ensuring maximum data compression and reliable RF beam predictions.

Figures (10)

• Figure 1: Beam alignment classical procedure, where $N_{\mathrm{BS}}$ is the number of antennas at BS and $O_{\mathrm{BS}}$ is the oversampling factor.
• Figure 2: System model illustration of the proposed autoencoder-based beam probing and hybrid precoding approach.
• Figure 3: The proposed autoencoder NN$_{\text{Auto-HP}}$ consists of two sections: the probing precoders generator (encoder NN$_{\text{enc}}$) and the RF precoders generator (decoder NN$_{\text{dec}}$), where $\mathbf{r}_{\mathrm{RSSI}}$, $\mathbf{y}_{\mathrm{RSSI}}$, $\mathbf{d}_{3}$, $\mathbf{d}_{2}$, $\mathbf{d}_{1}$ are the outputs of each hidden layer in NN$_{\text{Auto-HP}}$.
• Figure 4: Illustration of the Deployment mode.
• Figure 5: Illustration of the area covered in the DeepMIMO channel datasets by considered communication scenario.