Physical-Layer Jamming Signal Detection with Transfer Learning
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing jamming detection techniques in wireless communication networks face challenges due to heterogeneous device capabilities and limited data availability, leading to inaccurate model performance when using centralized or local training methods.
Innovation Solution
A method utilizing transfer learning to generate target ML models for jamming detection across nodes with different characteristics by training source ML models at a central node and deploying them using reserved resource elements, tailored to each node's specific features.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If centralized training of jamming detection models is performed, then model accuracy may be improved, but data transportation complexity and resource usage increase excessively
Solution Approach 1:
The patent segments the centralized training process into distributed training across multiple nodes. Each node trains local models using its own measurement data, eliminating the need to transport large volumes of IQ signal data to a central server. This segmentation maintains detection accuracy while significantly reducing data transportation complexity and resource usage.
Solution Approach 2:
The patent introduces a hierarchical dimension to the training architecture, combining local distributed training at the node level with selective centralized coordination. This dimensional approach allows nodes to independently train models locally (reducing data transport) while still benefiting from centralized model aggregation and knowledge sharing when appropriate, thus resolving the contradiction between accuracy and complexity.
2Device complexity
If independent models are trained on a per-node basis, then data transportation is reduced, but model accuracy decreases due to limited measurement data availability
Solution Approach 1:
The patent merges the strengths of both centralized and distributed approaches by implementing a hybrid architecture. Nodes combine their local measurement data to train models collectively, pooling resources across the network. This merging allows each node to benefit from aggregated data from multiple sources, improving model accuracy while maintaining the low data transportation complexity of distributed training.
Solution Approach 2:
The patent introduces a model aggregation mechanism as an intermediary between local training and centralized coordination. This intermediary collects trained models from multiple nodes, combines them through federated learning or model averaging, and distributes improved models back to nodes. This intermediary approach enables accuracy improvement through data pooling without requiring raw measurement data to be transported, thus resolving the contradiction.
3Use of energy by moving object
If AI/ML model training is performed at nodes with limited computing resources, then resource usage is reduced, but training time and complexity increase
Solution Approach 1:
The patent implements preliminary action by pre-training base models using aggregated data from multiple nodes before deployment. This preliminary training phase creates a robust foundation model that can be quickly adapted to specific node conditions through fine-tuning with minimal local data. This approach reduces both the computing resources needed at individual nodes and the overall training time, as the heavy lifting is done in advance.
Solution Approach 2:
The patent introduces dynamic adaptability to the training process, allowing nodes to adjust their training intensity and resource allocation based on available resources and urgency requirements. The system dynamically balances between quick local adaptation for immediate needs and more comprehensive centralized training when resources permit, thus optimizing the trade-off between resource usage and training time in varying operational conditions.
Data Source
AI summary
A method (200) for facilitating detection of jamming signals on the Physical layer of a communication network is disclosed. The method is performed by a first node of the communication network and comprises receiving, from each of a plurality of second nodes, a value of at least one parameter characterizing the respective second node (210), and obtaining measurement data of radio signals in a coverage area of the first node (220). The method further comprises determining, from at least the received parameter values, a number of source ML models to be trained to detect jamming signals (230), and training the determined number of source ML models using the obtained measurement data of radio signals in a coverage area of the first node (240). The method further comprises, for each trained source ML model, causing a target ML model to be generated from the source ML model using a Transfer Learning process (250), causing each of the plurality of second nodes to use at least one of the target ML models to detect jamming signals on the Physical layer of the communication network (260).