基于Stacking分类融合的海洋哺乳动物声学识别研究  

Acoustic identification of marine mammals based on Stacking classification fusion

作　　者：赵彪 蔡文郁[1,2] 祝嵇峰 ZHAO Biao;CAI Wenyu;ZHU Jifeng(School of Electronic Information,Hangzhou Dianzi University,Hangzhou Zhejiang 310018,China;Zhejiang Provincial Key Lab of Equipment Electronics,Hangzhou Zhejiang 310018,China)

机构地区：[1]杭州电子科技大学电子信息学院,浙江杭州310018 [2]浙江省装备电子研究重点实验室,浙江杭州310018

出　　处：《杭州电子科技大学学报（自然科学版）》2023年第4期7-13,共7页Journal of Hangzhou Dianzi University：Natural Sciences

基　　金：浙江省自然科学基金资助项目(LZ22F010004,LZJWY22E090001);浙江省属高校基本科研业务费专项资金资助项目(GK209907299001-001);浙江省装备电子研究重点实验室资助项目(2019E10009)。

摘　　要：在自主研制的高精度声学探测系统采集的海洋哺乳动物声学数据基础上,提出一种基于Stacking分类融合的声学数据处理算法。该方法融合了支持向量机(Support Vector Machines,SVM)、K最近邻(K-Nearest Neighbor,KNN)、决策树(Decision Tree)、朴素贝叶斯(Naive Bayes model)4种单分类模型,并采用Labview软件开发,将采集的水下哺乳动物音频数据提取梅尔频率倒谱(Mel-Frequency Cepstrum Coefficient,MFCC)系数作为分类的特征参数用于对海洋哺乳动物分类。采用6个海洋哺乳动物种类共4042个样本进行测试,和单分类模型识别率最高的SVM模型相比,该方法的识别率提升了3.30%,且在精准率、召回率、F_(1)值等分类评估指标中也有更好的表现。Based on the acoustic data of marine mammals collected by the self-developed high-precision acoustic detection system,an acoustic data processing algorithm based on Stacking classification fusion is proposed.This method combines four single classification models of Support Vector Machines(SVM),K-Nearest Neighbor(KNN),Decision Tree,and Naive Bayes model,and developed with Labview software;the Mel-Frequency Cepstrum Coefficient(MFCC)coefficients are extracted from the collected underwater mammal audio data as the characteristic parameters of the classification to classify marine mammals.A total of 4042 samples of 6 marine mammal species are used for testing.Compared with the SVM model with the highest recognition rate of single classification model,the recognition rate of this method is increased by 3.30%,and it also achieves better performance in the classifier evaluation metrics such as accuracy,recall rate and F_(1) value.

关 键 词：海洋哺乳动物 Stacking分类融合 MFCC 声学识别 

分 类 号：TN912.34[电子电信—通信与信息系统]