RAN Control Device with Distributed AI Retraining
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The existing RAN Intelligent Controllers (RICs) face challenges with increased processing load and delayed concept drift detection due to the concentration of AI/ML learning and retraining functions in Near-RT RICs, which lack sufficient computing resources and information from adjacent areas.
Innovation Solution
A control device is introduced that hierarchizes non-real time (Non-RT) and near-real time (Near-RT) RICs, distributing AI/ML learning and inference functions to Near-RT RICs and retraining functions to Non-RT RICs, with concept drift detection and retraining distributed across both units.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Speed
If AI/ML learning and retraining functions are concentrated in Near-RT RICs, then fast inference control is achieved, but processing load increases and computing resources become insufficient
Solution Approach 1:
The patent segments the AI/ML functions by separating the learning model generation and retraining functions from the inference function. The Near-RT RIC performs only inference using pre-generated learning models, while the Non-RT RIC performs the computationally intensive learning model generation and retraining. This segmentation reduces the processing load on Near-RT RICs while maintaining fast inference control capability.
2Reliability
If retraining is performed frequently to adapt to environmental changes, then model accuracy improves, but processing time and resource consumption increase
Solution Approach 1:
The patent implements preliminary action by performing learning model generation and retraining in advance in the Non-RT RIC, rather than waiting for concept drift to occur. The system monitors concept drift and triggers retraining proactively, allowing the learning model to be updated before performance degradation becomes significant. This approach maintains model accuracy while managing retraining time and resource consumption efficiently.
3Speed
If concept drift detection is performed using only local data, then detection speed is fast, but detection accuracy decreases due to lack of broader context
Solution Approach 1:
The patent merges local and non-local data sources for concept drift detection. The Near-RT RIC performs initial concept drift detection using local data for fast response, while the Non-RT RIC performs comprehensive analysis by combining local data with non-local data from other base stations. This combination provides broader context and improves detection accuracy while maintaining the speed advantage of local processing.
Data Source
AI summary
A non-real time control unit and a near-real time control unit are hierarchized, and a learning and inference unit, that controls the radio access network based on a result of inference performed by applying newest data of the data collected to a learning model generated based on data collected from an O-RAN base station device, is arranged in the near-real time control unit. Also, a part of a retraining unit, that detects concept drift based on a history of the data collected, and causes the learning and inference unit to retrain the learning model when the concept drift is detected, is arranged in the near-real time control unit, and other parts are placed in the non-real time control unit.