Road facility semantic topology lightweight graph neural network modeling method

By combining a decoupled architecture and a distillation loss function, the model redundancy problem caused by the coupling of topology and attribute features in existing technologies is solved, realizing lightweight road facility graph neural network modeling, which is suitable for edge computing devices.

CN122452373AActive Publication Date: 2026-07-24WUHAN WUDA ZOYON SCI & TECH
5 Cites 0 Cited by

Patent Information

Application Number
CN202610899819.9
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-06-22
Publication Date
2026-07-24
Estimated Expiration
2046-06-22

AI Technical Summary

Technical Problem

Existing technologies in neural network modeling of road infrastructure maps suffer from redundant model parameters and excessively high inference latency due to the feature coupling between topological distribution and attribute distribution, which fails to meet the requirements of lightweight and fast response for edge computing devices.

Method used

A decoupled architecture of topological feature extraction branch and semantic feature extraction branch is adopted. The graph Laplacian matrix is ​​constructed by the topological position coordinate sequence and first-order Chebyshev polynomial graph filtering is performed. The semantic attribute feature sequence is subjected to one-dimensional convolution operation. The network weights are updated by combining the distillation loss function of pre-trained large graph neural network to generate a lightweight model.

Benefits of technology

It reduces the number of nonlinear operations and floating-point operations in the model, lowers the number of parameters and inference response time, while maintaining the feature representation capability of complex semantic topological relationships, and adapts to the deployment requirements of edge computing devices.

✦ Generated by Eureka AI based on patent content.
Patent Text Reader

Abstract

The invention belongs to the technical field of electric digital data processing, and relates to a road facility semantic topology lightweight graph neural network modeling method. The method comprises the following steps: analyzing road facility data into a topological coordinate sequence and a semantic feature sequence; constructing a topology and semantic double-flow branch, constructing a graph Laplacian matrix in the topology branch according to a coordinate sequence, and performing graph filtering by adopting a first-order Chebyshev polynomial without a nonlinear activation function to extract a topology representation vector; stripping coordinate information in the semantic branches, and executing one-dimensional convolution on the semantic feature sequence to extract a semantic representation vector; splicing the two vectors to generate a fusion representation vector; a pre-trained large graph neural network of isomorphic topology is used as a teacher model, a mean square error of a teacher vector and a fusion representation vector is calculated to construct a distillation loss function, and the network weight is updated through back propagation to obtain a lightweight model. According to the method, the non-linear operation and floating point operation amount is reduced, the back propagation of topological noise is blocked, and the reasoning response time is shortened.
Need to check novelty before this filing date? Find Prior Art