Nonlinear Signal Dimension Reduction for 5G Fronthaul Compression
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The existing linear transforms used in radio units with large antenna arrays, such as SVD and DFT, fail to provide efficient signal concentration and dimension reduction, leading to high front-haul data rates and computational load in baseband units, especially in 5G networks with large antenna arrays.
Innovation Solution
Implementing a non-linear transformation using an alternating sequence of linear and nonlinear functions, specifically a trained convolutional neural network autoencoder, to reduce signal dimensions and improve sparse representations, allowing for more efficient communication between radio and baseband units.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of manufacture
If linear transforms (SVD, DFT) are used for signal dimension reduction, then the transformation is computationally simple, but signal concentration is insufficient leading to high front-haul data rates
Solution Approach 1:
The patent changes the transformation type from linear to non-linear, using a trained neural network model that adapts its parameters (weights and biases) based on training data. This allows the system to learn optimal signal representations that concentrate energy more effectively than fixed linear transforms, reducing the number of bits needed for front-haul transmission while maintaining signal quality.
Solution Approach 2:
The patent replaces traditional mathematical transform mechanisms (SVD, DFT) with a data-driven neural network mechanism. The neural network learns complex non-linear mappings from training data, substituting the rigid mechanical-like operations of linear algebra with flexible, adaptive computational patterns that achieve better signal concentration.
2Device complexity
If linear transforms are used for dimension reduction, then device complexity is low, but signal concentration and sparse representation are insufficient
Solution Approach 1:
The patent performs preliminary training of the neural network offline using representative signal data. This preliminary action allows the network to learn optimal transformation patterns before actual signal processing, enabling it to achieve superior signal concentration and sparse representation during operation without adding real-time computational complexity.
Solution Approach 2:
The patent transforms the fixed parameters of linear transforms into learnable parameters of a neural network. By changing from static transformation matrices to adaptive neural network weights, the system achieves better signal concentration quality while managing complexity through efficient network architecture and pre-training.
3Reliability
If more antenna elements are used in AAS, then beamforming capability is improved, but computational load in baseband unit increases
Solution Approach 1:
The patent extracts only the most significant signal components after neural network transformation, transmitting a reduced set of bits to the baseband unit. This extraction principle allows the system to maintain beamforming capability with many antenna elements while reducing the computational load at the baseband unit by processing only the essential signal information.
Solution Approach 2:
The patent changes the representation parameters of the antenna signals through non-linear transformation, converting high-dimensional antenna element data into a compact, concentrated form. This parameter transformation reduces the amount of data requiring baseband processing while preserving the beamforming capabilities enabled by multiple antenna elements.
4Quantity of substance
If dimension reduction is applied to reduce front-haul data rate, then transmission requirements are reduced, but signal information loss may increase
Solution Approach 1:
The patent applies non-linear parameter transformation through a trained neural network that learns to preserve essential signal information while compressing the representation. The network adjusts its parameters during training to minimize information loss, achieving effective dimension reduction that maintains signal fidelity and minimizes information loss in the compressed domain.
Data Source
Figure 1
Figure 2
Figure 3~4
AI summary
Embodiments herein e.g. discloses a method performed by a radio unit (13) for handling a number of received radio signals over an array of antennas comprised in the radio unit (13). The radio unit transforms the number of received radio signals into a number of sequences of complex symbols. The radio unit further filters the number of sequences of complex symbols by inputting the number of sequences of complex symbols into a trained computational model comprising an alternating sequence of linear and nonlinear functions and thereby obtaining a reduced number of sequences. The radio unit further transmits the reduced 10number of sequences to a baseband unit (12) over a front-haul link.