Radio IQ signal-oriented base model distributed pre-training method

CN121940251AActive Publication Date: 2026-04-28ARTIFICIAL INTELLIGENCE INNOVATION RES INST OF ZHEJIANG UNIV OF TECH BINJIANG DISTRICT HANGZHOU
View PDF 3 Cites 0 Cited by

Patent Information

Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
ARTIFICIAL INTELLIGENCE INNOVATION RES INST OF ZHEJIANG UNIV OF TECH BINJIANG DISTRICT HANGZHOU
Filing Date
2026-03-30
Publication Date
2026-04-28

AI Technical Summary

Technical Problem

Existing technologies in radio signal processing suffer from problems such as data heterogeneity, scarcity of labeled data, lack of unified standardization processes, low model training efficiency, poor training stability, and unbalanced computational load in distributed training, making it difficult to achieve generalization and transfer capabilities across tasks and scenarios.

Method used

A distributed pre-training method for pedestal models oriented towards radio IQ signals is adopted. By power normalization and channel-independent segmentation embedding, combined with self-attention modeling and cross-channel attention interaction, the batch size is dynamically adjusted to perform joint training on multiple datasets. The model is optimized by mask reconstruction loss and finally applied to fine-tuning for downstream tasks.

Benefits of technology

It improves the discriminative power of signal representation and the generalization ability of the model, enhances the utilization rate of GPU memory and training stability of multi-source heterogeneous data, and can quickly adapt to various downstream radio signal processing tasks while maintaining robust recognition performance.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN121940251A_ABST
    Figure CN121940251A_ABST
Patent Text Reader

Abstract

The invention discloses a base model distributed pre-training method for radio IQ signals, and belongs to the technical field of signal processing and artificial intelligence. Aiming at the problems that radio signal data are heterogeneous, modeling is difficult to consider channel independence and interaction, pre-training alignment is difficult and distributed training efficiency is low, the method comprises the following steps: performing power normalization and channel independent segmentation embedding on I / Q signals; feature coding is carried out through a two-stage mechanism of self-attention and cross-channel attention in the channel; performing reconstruction pre-training by adopting a channel independent random mask; based on the dynamic batch size, multi-data-set sampling and explicit fragmentation, efficient distributed training is achieved; and finely tuning the pre-training model for downstream tasks. The method can improve signal characterization quality and training efficiency, and is suitable for tasks such as modulation recognition, individual recognition and anomaly detection.
Need to check novelty before this filing date? Find Prior Art

Citation Information

Patent Citations

  • Data processing method, electronic equipment and computer readable storage medium

    CN120929906A

  • Multimodal financial technology deep learning core with joint optimization of vector-quantized variational autoencoder and neural upsampler

    US12327190B1

  • Non-autoregressive transformer-based modeling method for 4-level pulse amplitude modulation high-speed transmitter

    WO2026020649A1