AI Model Knowledge Distillation for Base Station Traffic Prediction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing machine learning-based methods for traffic prediction in communication networks face challenges due to changing traffic patterns and limited data storage capacity at base stations, leading to accuracy degradation and high costs associated with data collection and storage.
Innovation Solution
A server-based system that uses knowledge distillation to update AI models for predicting communication loads at base stations, where the system obtains traffic data sets, trains AI models to extract features and predict loads, and updates models by converging aggregated losses, allowing for efficient prediction without relying on extensive historical data.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional ML or NN models are trained on large amounts of historical data to improve prediction accuracy, then prediction accuracy is improved, but data storage capacity requirements increase and operational costs increase
Solution Approach 1:
The patent extracts only the essential knowledge and patterns from historical data through AI model training, rather than storing the raw historical data itself. The AI model parameters and learned representations are retained as compressed knowledge, enabling accurate predictions without maintaining large volumes of original traffic data at base stations.
Solution Approach 2:
The patent creates a simplified copy of the prediction capability through AI models that can be deployed at base stations. Instead of copying and storing extensive historical data, the system trains AI models on centralized servers and deploys these model copies to edge devices, enabling local predictions with minimal data storage requirements.
2Adaptability or versatility
If traffic patterns change in the real world, then adaptability to new situations is improved, but prediction accuracy degrades because models trained on previous data cannot generalize effectively
Solution Approach 1:
The patent implements dynamic model updating where AI models are continuously retrained and updated as new traffic data becomes available. The system adapts to changing traffic patterns by periodically retraining models on recent data and deploying updated model versions, ensuring predictions remain accurate despite evolving network conditions.
Solution Approach 2:
The system incorporates feedback mechanisms where prediction performance is continuously monitored and used to trigger model retraining when accuracy degradation is detected. This feedback loop enables the system to automatically adapt to changing traffic patterns by identifying when retraining is needed and initiating the update process.
3Measurement precision
If data is collected from multiple base stations to improve prediction accuracy, then prediction accuracy is improved, but bandwidth cost increases and computing resource utilization degrades
Solution Approach 1:
The patent segments the prediction system into centralized training and distributed inference components. Data collection and model training are performed centrally on servers with sufficient resources, while base stations perform only lightweight inference using deployed models. This segmentation eliminates the need for continuous data migration and frequent disk writes at edge devices.
Solution Approach 2:
The patent introduces an intermediary centralized server that aggregates training data from multiple base stations and performs model training. This intermediary handles the computationally intensive tasks of data processing and model retraining, while base stations simply provide data samples and receive updated models, significantly reducing bandwidth consumption and computing overhead at the network edge.
Data Source
AI summary
A server for predicting a future traffic load of a base station is provided. The server may obtain a first prediction model based on traffic data collected from the base station for a first period of time, obtain a second prediction model based on traffic data collected from the same base station for a second period time, and also based on knowledge transferred from the first prediction model. Each of the first prediction model and the second prediction model may include an encoder module, a reconstruction module, and a prediction module which are connected to form two paths, an encoder-reconstruction path and an encoder-prediction path, to preserve more information of historic traffic data.