基于多尺度分频和变权的战场声目标识别

Acoustic Target Recognition in Battlefield Based on Multiscale Frequency Division and Nonuniform Weight

提出一种基于多尺度分频和变权的战场声目标分类与识别的快速算法,为战场环境下快速识别典型声目标提供了一种实用方法。首先对信号进行离散傅立叶变换,获得幅频谱,对幅频谱进行平滑递推滤波,之后对其进行多尺度分频,在同一尺度上求取幅值均值,并对其进行定量归一化,将归一化后的数组作为此尺度上的特征向量,多次训练获得大量特征向量数组,基于相似性系数的特性,从中提取特征向量模板,并给出模板相应的权重。最后,采用基于相似性系数的分类器,对信号进行分类与识别。并将此算法与传统的特征提取算法进行试验对比。试验结果表明,基于多尺度分频和变权的战场声目标分类与识别算法简单有效、识别率高、计算周期短、实时性好、适合目标跟踪时间较短的场合。

A fast algorithm of classification and recognition of acoustic target in battlefield based on multiscale frequency division and nonuniform weight was proposed, which provided a utility method for fast recognition of typical target in battlefield. After performing fast Flourier transform （FFT） to the signal, the amplitude-frequency spectrum chart of the signal was obtained. The amplitude-frequency spectrum was treated by smoothing recursive filter and multiscale frequency division. The amplitude mean was calculated in the same scale. The eigenvector is the quantitative normalized frequency mean in the same scale. The eigenvector templates were extracted from eigenvector array which was composed of eigenvectors by training different signals of the same target based on similarity coefficient. The weights corresponding to the eigenvector templates were given. The separator was designed based on similarity coefficient, which was used for the classification and recognition of acoustic signals. The algorithm was compared with the traditional one. The hardware-in-the-loop simulation represents that the algorithm of acoustic target classification and recognition in battlefield based on multiscale frequency division and nonuniform weight is fast and effective, with high recognition rate, fine real-time responsibility and short calculation time, which can be used where the tracking time is short.