凡纳滨对虾摄食不同饵料的声音信号分类模型研究  

Study on classification models for acoustic signals of Litopenaeus vannamei feeding on different kinds of diets

作　　者：曹正良 王子贤 李钊丞 姜珊 蒋千庆 胡庆松[3] 靳雨雪 CAO Zhengliang;WANG Zixian;LI Zhaocheng;JIANG Shan;JIANG Qianqing;HU Qingsong;JIN Yuxue(College of Oceanography and Ecological Science,Shanghai Ocean University,Shanghai 201306,China;College of Marine Living Resource Sciences and Management,Shanghai Ocean University,Shanghai 201306,China;College of Engineering Science and Technology,Shanghai Ocean University,Shanghai 201306,China)

机构地区：[1]上海海洋大学海洋科学与生态环境学院,上海201306 [2]上海海洋大学海洋生物资源与管理学院,上海201306 [3]上海海洋大学工程学院,上海201306

出　　处：《南方水产科学》2025年第2期27-37,共11页South China Fisheries Science

基　　金：上海市农业科技创新项目(T2023108)。

摘　　要：利用机器学习技术对凡纳滨对虾(Litopenaeus vannamei)摄食不同饵料时的声音信号进行分类,旨在比较不同分类模型的性能,确定最优模型,为对虾养殖中饵料管理的信息化提供参考。研究选取对虾摄食沙蚕、颗粒饲料和鱿鱼时的声音信号,经降噪滤波处理后,通过2类方式分类:1)基于音频特征向量,分别建立支持向量机(Support Vector Machine,SVM)、随机森林(Random forest,RF)和k-最近邻(k-Nearest Neighbor,KNN)模型;2)基于梅尔频谱图,建立卷积神经网络(Convolutional Neural Networks,CNN)模型。结果表明,结合Mixup数据增强技术和粒子群优化算法(Particle Swarm Optimization,PSO)的CNN模型在准确率方面表现最佳,达到91.67%。4个模型在识别颗粒饲料的召回率均超过90%,说明摄食颗粒饲料的声音信号相较于摄食沙蚕和鱿鱼等软体饵料更易识别。CNN模型不仅在准确率、精确度、召回率和F1分数等指标上均优于上述传统模型,同时能够适应复杂声学信号的分析需求,具有较大的应用潜力。We explored the classification of acoustic signals of L.vannamei feeding on different kinds of diets with machine learning techniques,so as to recognize the best model by comparing the performance among different classification models,and to provide references for the informatisation of feed management in shrimp aquaculture.We selected and processed the acous-tic signals of L.vannamei feeding on nereid,pellet feed and squid.After noise reduction and filtering,two classification ways were employed:1)Building models including Support Vector Machine(SVM),Random Forest(RF),and k-Nearest Neighbor(KNN)based on audio feature vectors.2)Building Convolutional Neural Network(CNN)based on Mel-spectrograms.The results indicate that the CNN model,enhanced with Mixup data augmentation and Particle Swarm Optimization(PSO),achieved the highest accuracy of 91.67%.In addition,all four models achieved a recall rate exceeding 90%in identifying pellets,which indicates that the acoustic signals of shrimps feeding on pellets were more distinguishable than those associated with nereid and squid.The CNN model outperformed the traditional models in accuracy,precision,recall,and F1 score,exhibiting greater adaptability for analyzing complex acoustic signals with significant potential for its practical application.

关 键 词：凡纳滨对虾 被动声学 声音信号分类 机器学习 卷积神经网络模型 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程] TB561[自动化与计算机技术—控制科学与工程] S966[交通运输工程—水声工程]