Radio IQ signal-oriented base model distributed pre-training method
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
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.
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.
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.
Smart Images

Figure CN121940251A_ABST
Abstract
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