Multi-modal data alignment and enhancement method and device in industrial large model training

By constructing dynamic anchor points and industrial knowledge graphs during the training of large industrial models to align multimodal data, and using adversarial generative networks to generate data that conforms to physical laws, the challenges in multimodal data alignment and enhancement are solved, thereby improving the training efficiency and decision credibility of the model.

CN122332948APending Publication Date: 2026-07-03BEIJING EASY TIMES DIGITAL TECH
0 Cites 1 Cited by

Patent Information

Application Number
CN202610307789.8
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-03-13
Publication Date
2026-07-03

AI Technical Summary

Technical Problem

In the training of large industrial models, multimodal data are difficult to align precisely on the time axis, cross-modal semantic associations are fragmented, and traditional data augmentation methods generate data with high distortion rates, which cannot meet the credibility requirements of model decisions in safety-critical scenarios.

Method used

By detecting physical events to determine dynamic anchor points, constructing a local time scale for spatiotemporal alignment, and performing feature mapping and semantic alignment under the constraints of an industrial knowledge graph, combined with an adversarial generative network, high-quality training data that satisfies statistical characteristics and physical laws is generated using a realism and physical consistency discriminator.

Benefits of technology

It significantly improves the spatiotemporal alignment accuracy and semantic expression consistency of multimodal data, and the generated data conforms to physical reality. It improves the training efficiency, anomaly detection accuracy and decision credibility of the model, and meets the high precision and high real-time requirements of industrial scenarios.

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

Abstract

According to the multi-modal data alignment and enhancement method and device in industrial large model training provided by the embodiment of the invention, heterogeneous features such as vibration, vision and process parameters are aligned in a semantic space conforming to an engineering law through a cross-modal semantic distillation framework which fuses an industrial knowledge graph and physical law constraints; the problem of semantic expression splitting is effectively solved, and the understanding depth of the industrial process is enhanced; by introducing a double-discriminator adversarial enhancement strategy comprising an authenticity discriminator and a physical consistency discriminator and combining with the constraint of an industrial abnormal mode library, the generated data is ensured to meet statistical distribution characteristics and conform to physical reality, and the data distortion rate is greatly reduced, so that high-fidelity training data is provided for an industrial large model, and the training efficiency is improved. And finally, the model convergence speed, the anomaly detection precision and the decision credibility are comprehensively improved, and the application requirements of an industrial scene for high precision and high real-time performance are met.
Need to check novelty before this filing date? Find Prior Art