EI / SCOPUS / CSCD 收录

中文核心期刊

基于多通道多尺度卷积注意力网络的次声事件分类方法

Infrasound event classification with multi-channel multi-scale convolutional attention network

  • 摘要: 次声波在传播过程中易受大气多径效应影响, 导致信号混叠和干扰, 从而影响分类性能, 为此提出了一种基于多通道多尺度卷积注意力网络的次声信号分类方法。首先, 采用完全自适应噪声集合经验模态分解方法分解原始信号, 将其中的四个固有模态函数进行维度拼接, 得到多通道特征组合, 降低多径分量与直达信号的混叠程度; 其次, 构建多尺度卷积注意力网络, 通过多尺度卷积核同时捕获特征的长时相关性和局部突变特性, 并利用通道−空间注意力机制自适应聚焦关键信息。实验结果表明, 该方法能显著提升多径环境下的分类准确率, 在次声数据集上取得了82.76%的平均分类精度, 优于其他两类传统次声信号分类方法, 证明了该方法在次声信号分类中的有效性和鲁棒性。

     

    Abstract: Infrasound signals are prone to atmospheric multipath effects during propagation, resulting in signal aliasing and interference that adversely affect classification performance. To overcome this challenge, a method for classifying infrasound signals based on multi-channel multi-scale convolutional attention network is proposed. The proposed approach first employs the complementary ensemble empirical mode decomposition with adaptive noise to decompose the original signal, then performs dimensional concatenation of four intrinsic mode functions along the feature axis to construct a multi-channel feature representation, effectively mitigating aliasing between multipath components and direct signals. Subsequently, a multi-scale convolutional attention network is constructed, utilizing multi-scale kernels to simultaneously capture both long-term dependencies and local transient features, with a channel-spatial attention mechanism adaptively emphasizing discriminative information. Experimental results demonstrate the method’s superior performance in multipath environments, achieving 82.76% average classification accuracy on the infrasound dataset and outperforming two conventional classification approaches, thereby confirming its effectiveness and robustness for infrasound signal classification.

     

/

返回文章
返回