User Plane Data Exchange for Federated RAN AI Model Updates
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing communication networks lack efficient data collection mechanisms when the Xn interface is unavailable, have a limited scope of distributed and federated learning, decouple AI/ML applications in the RAN from the application layer, and struggle with hyperparameter tuning and model updates based on metrics like prediction accuracy and F1 score.
Innovation Solution
Implement a system that facilitates distributed and federated learning by using enhanced containers and protocols to exchange AI-related data between the core network and RAN, enabling training and updating of global and local AI models, even without an Xn interface, and leveraging the core network's infrastructure for AI/ML model training and updates.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If distributed and federated learning is implemented across RAN nodes and core network, then AI model accuracy and network performance are improved, but device complexity and data exchange requirements increase
Solution Approach 1:
The AI model training process is segmented into distributed components across multiple RAN nodes and core network entities. Each node trains local models independently using local data, then these local models are aggregated in the core network to form a global model. This segmentation allows parallel processing and reduces the computational burden on any single node while improving overall model accuracy through diverse data sources.
Solution Approach 2:
The patent implements a nested architecture where local AI models at RAN nodes are embedded within the broader federated learning framework of the core network. The local models contain specialized knowledge for specific nodes, while the global model in the core network contains the aggregated knowledge. This nested structure allows each level to operate independently while contributing to the overall system intelligence.
2Productivity
If AI-related data is exchanged between core network and RAN nodes, then model training capability is enhanced, but network bandwidth consumption and data transmission requirements increase
Solution Approach 1:
The patent extracts only the essential model parameters and gradient information needed for training from each RAN node, rather than transmitting complete datasets or full model states. By taking out only the critical training components (model weights, gradients, and metadata), the system achieves effective distributed training while minimizing network bandwidth consumption and data transmission requirements.
3Speed
If local AI models are trained at individual base stations, then response time and local decision-making capability are improved, but model consistency and coordination across nodes become more difficult
Solution Approach 1:
The patent implements a feedback mechanism where local AI models at RAN nodes continuously transmit their training results, performance metrics, and model parameters to the core network. The core network analyzes this feedback and generates updated global model configurations that are then distributed back to all nodes. This closed-loop feedback system ensures that local models remain consistent with the global model while maintaining their ability to respond quickly to local conditions.
Data Source
AI summary
Management and performance of distributed and federated learning can be enhanced. Core network can comprise UPF, AF, and AI component. AI component can train global AI model, located in core network, based on first AI-related data contained in a first container received from first base station by UPF. The trained global AI model can generate second AI-related data based on first input data input to the trained global AI model. AF and/or UPF can communicate a second container, comprising the second AI-related data, to second base station to facilitate training or updating a local AI model located at second base station. The trained or updated local AI model can generate a prediction or inference based on second input data input to the trained or updated local AI model. Second base station can communicate information relating to the prediction or inference to a device associated with second base station.