Material whole-process tracing method based on federated learning and knowledge graph

Abstract

The present application belongs to the field of material life cycle data management, and discloses a material whole-process tracing method based on federated learning and knowledge graph, which is used to solve the problems of long link, scattered data sources, difficulty in unified identification and association, missing due to unstable communication in the south pole, and inability to centralized modeling due to data barriers of multiple units, and lack of quantitative maintenance and replacement basis. The present application generates a unified material identification, constructs a local time sequence knowledge graph containing a time stamp and a confidence degree, obtains a material state embedding through graph embedding with confidence weighting and time decay, performs generative missing completion with the embedding as a condition, quantifies the uncertainty, and writes back the graph confidence degree, trains a global model through asynchronous federated aggregation with delay compensation and credibility weighting, realizes cross-subject safe collaborative modeling, traces link credibility quantization and life risk assessment, and outputs maintenance and replacement suggestions.

Images