Neural Network Uplink RAN Pre-Processing for Fronthaul Compression
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
In 5G MIMO systems, the radio head apparatus and central processing apparatus face challenges in efficiently compressing high-dimensional signals over fronthaul due to varying channel conditions, leading to increased processing costs and potential degradation, especially when using uniform quantization schemes.
Innovation Solution
Implementing neural networks at the radio head apparatus for data signal pre-processing, selecting an appropriate neural network based on channel information, and jointly training with the central processing apparatus to optimize compression and quantization, while preserving signal quality and reducing dimensionality.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If known quantization schemes (uniform quantization, Lloyd-Max quantization, or Grassmanian manifold based quantization) are used to directly treat the received signal at the radio head apparatus, then quantization can be performed, but severe degradation occurs due to the large dimension of the received signal
Solution Approach 1:
The patent divides the quantization process into two stages: first, a compression network compresses the high-dimensional received signal into a lower-dimensional representation; second, the compressed signal is quantized using traditional quantization schemes. This segmentation allows traditional quantization to work effectively on reduced-dimensional data while the compression network handles the dimensionality reduction challenge.
Solution Approach 2:
The compression network performs preliminary compression of the received signal before quantization is applied. By pre-compressing the signal to reduce its dimensionality, the system prepares the signal in a form that is suitable for subsequent quantization operations, avoiding the severe degradation that would occur if quantization were applied directly to the high-dimensional signal.
2Productivity
If more functions are performed at the radio head apparatus, then there is less strain on the fronthaul, but this comes at an increased cost in terms of processing and memory capabilities at the radio head apparatus side
Solution Approach 1:
The patent changes the parameter of signal dimensionality through the compression network. By transforming the high-dimensional signal into a lower-dimensional representation, the system reduces the amount of data that needs to be processed and transmitted, thereby reducing the processing and memory requirements at the radio head apparatus while still performing advanced signal processing functions.
3Ease of manufacture
If more functions are put on the central processing apparatus side, then costs are lower, but the fronthaul capacity becomes a bottleneck
Solution Approach 1:
The patent segments the signal processing functions between the radio head apparatus and the central processing apparatus. The compression network at the radio head performs dimensionality reduction locally, sending only the compressed lower-dimensional signal over the fronthaul to the central processing apparatus for quantization and further processing. This segmentation reduces fronthaul data volume while distributing computational tasks appropriately.
4Measurement precision
If neural networks are used for compression at the radio head apparatus, then signal dimensionality is reduced and quantization noise is mitigated, but processing complexity at the radio head increases
Solution Approach 1:
The compression network changes the dimensional parameter of the signal, transforming high-dimensional input into lower-dimensional output. This parameter transformation reduces both the signal dimensionality and the associated quantization noise, while the network architecture is designed to achieve this compression efficiently.
Data Source
Figure 1~2
Figure 3
Figure 4
AI summary
In a base station of a radio access network, a distribution unit is confïgured to receive, through a radio head apparatus of the base station, a channel information signal transmitted by a user equipment over a radio channel, obtain based on the channel information signal compression model information indicating a neural network to be used for compression by the radio head apparatus amongst a set of neural networks, and sending the compression model information to the radio head apparatus. The radio head apparatus is confïgured to receive the compression model information from the central processing apparatus, receive a data signal from a user equipment over a radio channel, pre-proces the data signal, including compress the data signal by using a neural network, the neural network being selected based on the compression model information sent by the distribution unit, and transmit the pre-processed data signal to the central processing apparatus.