A training method for a multi-source direction-of-arrival (DOA) estimation model, a multi-source DOA estimation method, equipment, medium, and product.

By improving the position encoding and axial attention mechanism of the Transformer model and combining it with a multi-head attention mechanism, the structure is optimized into a pure encoder, which solves the problems of resolution and computational efficiency in direction-of-arrival estimation under multi-source and low signal-to-noise ratio environments, and achieves high-precision direction-of-arrival estimation for multi-source sound sources.

CN120910566BActive Publication Date: 2026-03-13NANJING UNIV OF AERONAUTICS & ASTRONAUTICS
View PDF 2 Cites 0 Cited by

Patent Information

Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2025-08-15
Publication Date
2026-03-13

AI Technical Summary

Technical Problem

Existing direction-of-arrival estimation methods have limited resolution, low computational efficiency, and poor accuracy in complex environments such as multiple sound sources and low signal-to-noise ratio, making them difficult to adapt to complex noise interference and multi-source interactions in real-world scenarios.

Method used

An improved Transformer model is adopted, which captures complex dependencies between long time periods and multiple frequency bands through a position encoding module and an axial attention mechanism, combined with a multi-head attention mechanism. It is optimized into a pure encoder structure, and a multi-task output module is introduced to reduce computational complexity and improve the ability to distinguish sound sources.

Benefits of technology

This technology enables high-precision direction-of-arrival estimation in multi-source, low signal-to-noise ratio environments, reducing computational costs, improving resolution, and adapting to complex sound source scenarios. It is applicable to array signal processing, sound source localization, radar, sonar, and wireless communication.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN120910566B_ABST
    Figure CN120910566B_ABST
Patent Text Reader

Abstract

This application discloses a training method, device, medium, and product for a multi-source direction-of-arrival (DOA) estimation model, relating to the interdisciplinary fields of signal processing and artificial intelligence. The training method includes: calculating the time-frequency features and cross-correlation features of the original signal from a multi-channel array; performing position encoding on the time-frequency features and microphone position information respectively; fusing these features into an input matrix and inputting it into the backbone network of an improved Transformer model; using a first multi-head self-attention layer to calculate the attention score of the input matrix and generate head output; then inputting the head output into a second multi-head self-attention layer to calculate the attention score and generate head output; finally, inputting the output into a multi-task output module to obtain the direction estimation result. This application introduces a multi-head attention mechanism to capture complex dependencies between long time periods and multiple frequency bands, improving the ability to distinguish sound sources. Multiple heads capture directional information from different perspectives, ensuring high accuracy even in complex environments.
Need to check novelty before this filing date? Find Prior Art

Citation Information

Patent Citations

  • Radio frequency fingerprint identification method and system based on convolution-attention mechanism and multi-packet reasoning

    CN120337998A

  • Temporal alignment of signals using attention

    WO2023219751A1