ML Beam Prediction Model Selection for Wireless Networks
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current beam management techniques in wireless communication systems, such as those used in LTE and 5G NR, are inefficient due to the time-consuming and resource-intensive process of sweeping all beams for beam selection and prediction, leading to increased latency and signaling overhead, which can reduce system throughput and spectral efficiency.
Innovation Solution
Implementing a method where user devices and network nodes use machine learning models to dynamically adjust the quantity of reference signal measurements for beam prediction inference, allowing for the selection of appropriate machine learning or non-machine learning algorithms based on beam change dynamics, enabling concurrent use of multiple models for improved efficiency and accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If beam sweeping is performed for all beams to ensure accurate beam selection and prediction, then beam selection accuracy is improved, but measurement overhead and latency increase
Solution Approach 1:
The patent applies partial action by performing beam sweeping only on a subset of beams rather than all beams. The network device determines a first quantity of reference signal measurements that is less than the total number of beams, performs sweeping only on these selected beams, and uses machine learning models to predict the remaining beams. This reduces measurement overhead and latency while maintaining acceptable beam selection accuracy through the predictive capability of the machine learning models.
2Measurement precision
If beam sweeping is performed for all beams to ensure accurate beam selection and prediction, then beam selection accuracy is improved, but signaling overhead increases
Solution Approach 1:
The patent reduces signaling overhead by performing beam sweeping only on a first quantity of reference signals that is less than all beams. The network device determines this reduced quantity, performs sweeping selectively, and supplements the results using machine learning-based beam prediction for the remaining beams. This partial sweeping approach significantly reduces the amount of signaling required while maintaining beam selection accuracy through the predictive models.
3Loss of time
If machine learning models use fewer reference signal measurements to reduce overhead, then measurement overhead is reduced, but beam prediction accuracy may deteriorate
Solution Approach 1:
The patent introduces machine learning models as an intermediary between the reduced reference signal measurements and the final beam prediction. The network device performs beam sweeping on only a first quantity of reference signals (fewer than all beams), then uses the machine learning model to infer beam predictions for the remaining beams based on the limited measurement data. This intermediary ML approach enables accurate beam prediction despite using fewer measurements, resolving the trade-off between overhead reduction and accuracy maintenance.
4Device complexity
If a single machine learning model is used for beam prediction, then device complexity is reduced, but adaptability to different beam dynamics conditions deteriorates
Solution Approach 1:
The patent implements dynamic model selection where the network device monitors beam change dynamics events and adaptively switches between different machine learning models based on the current beam dynamics conditions. When beam dynamics are detected to have changed, the network device selects an appropriate ML model from multiple available models that is best suited for the new conditions. This dynamic adaptation enables the system to handle varying beam dynamics effectively while maintaining manageable complexity through automated model selection rather than requiring manual configuration of multiple models.
Data Source
AI summary
A method includes beam prediction inference using a first machine learning (ML) model based algorithm that uses a first quantity of reference signal (RS) measurements; transmitting an indication that the first user device is using the first ML model based algorithm that uses the first quantity of RS measurements; receiving, based on beam change dynamics event information received by the network node from one or more other user devices that have one or more corresponding conditions within a threshold to the first user device, a request for the first user device to either: change to a non-ML model based algorithm to perform beam selection or a second ML model based algorithm that uses a second quantity of reference signal measurements to perform beam prediction inference, wherein the second quantity is different than the first quantity, or concurrently perform beam prediction inferences using the first and second ML model based algorithms.