基于时频域特征挖掘与自注意力机制融合的雷达PRI变化类型识别

Radar PRI variation mode recognition based on time-frequency domain feature mining and fusion via self-attention mechanism

摘要:

针对存在异常值时雷达辐射源脉冲重复间隔(pulse repetition interval, PRI)变化类型识别困难的问题,提出一种基于时频域特征挖掘与自注意力机制融合的雷达PRI变化类型识别方法。首先对PRI序列进行时序变化特征和小波特征分析,从时域和频域2个角度构建特征集;然后基于自注意力机制以数据驱动的方式学习时频特征之间的互补性,有效把握不同维度特征对识别效果的贡献,实现对不同维度特征的深度融合;最后基于全连接神经网络对融合后的特征进行模式分类,从而实现对PRI变化类型的识别。仿真结果表明,在不同异常值水平下,所提方法能够显著提高对6种典型PRI变化类型的识别准确率,而且识别效果要显著优于仅使用单一维度特征的方法。

Abstract:

Addressing the difficulty in recognising variation mode of the pulse repetition interval (PRI) of radar emitter when outliers are present, a method for recognising variation mode of PRI based on time-frequency domain feature mining and fusion via self-attention mechanism was proposed. Firstly, the time-varying characteristics and wavelet features of the PRI sequence were analyzed, and a feature set was constructed from both the time domain and frequency domain perspectives; then, based on the self-attention mechanism, it learned the complementarity between time-frequency features in a data-driven manner, effectively grasped the contribution of features from different dimensions to the recognition effect, and achieved deep fusion of features from different dimensions; finally, based on the fully connected neural network, the fused features were classified into patterns to achieve the recognition of PRI variation mode. Simulation results indicate that under different levels of outliers, the proposed method can significantly improve the recognition accuracy for 6 typical PRI variation mode. Moreover, its recognition performance is substantially superior to methods that only utilize single-dimensional features.