一种面向智能驾驶场景的轻量化参数共识聚合方法及系统
By decomposing the weight matrix of the pre-trained large language model and aggregating the dynamic parameters consensus, the problems of personalized adaptation and resource constraints of vehicle models in intelligent driving scenarios are solved, achieving efficient optimization and accuracy improvement under non-independent and identically distributed data.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- UNIV OF SCI & TECH BEIJING
- Filing Date
- 2025-10-28
- Publication Date
- 2026-07-17
AI Technical Summary
Existing technologies cannot effectively solve the problems of personalized adaptation of large vehicle models, non-independent and identically distributed data, and limitations in computing and communication resources in intelligent driving scenarios, resulting in slow model convergence and decreased accuracy.
By decomposing the weight matrix of the pre-trained large language model, freezing the general knowledge subspace, and using local data for gradient updates and parameter importance selection, sparse parameter increments are generated. These increments are then dynamically fused and averaged with the consensus parameter increments and selection matrix from the roadside server to achieve lightweight parameter consensus aggregation.
While protecting privacy, the optimization stability and accuracy of the vehicle model have been improved, adapting to non-ideal data distributions, reducing computational and communication overhead, and enhancing the robustness of the model in resource-constrained and dynamic scenarios.
Smart Images
Figure CN121505851B_ABST
Abstract
Citation Information
Patent Citations
Vehicle-mounted sensing equipment joint learning method for model structure optimization under edge computing
CN113595993A
Unmanned driving dynamic path planning method and system based on multi-source data fusion
CN120552911A