基于信号序列优化的蜂群状态精准识别机器听觉模型  

Accurate detection of bee colony status based on signal sequence optimized machine auditory model

作　　者：叶大鹏[1,2,3] 陈林杰 张林通 张雯清 魏增辉 黄少康[4] 瞿芳芳 YE Dapeng;CHEN Linjie;ZHANG Lintong;ZHANG Wenqing;WEI Zenghui;HUANG Shaokang;QU Fangfang(College of Mechanical and Electrical Engineering;College of Future Technology;Center for Artificial Intelligence in Agriculture;College of Bee Science and Biomedicine,Fujian Agriculture and Forestry University,Fuzhou,Fujian 350002,China)

机构地区：[1]福建农林大学机电工程学院 [2]福建农林大学未来技术学院 [3]福建农林大学农业人工智能研究中心 [4]福建农林大学蜂学与生物医药学院,福建福州350002

出　　处：《福建农林大学学报(自然科学版)》2025年第2期268-278,共11页Journal of Fujian Agriculture and Forestry University:Natural Science Edition

基　　金：福建省农业信息感知技术重点实验室建设补助项目(KJG22052A)。

摘　　要：【目的】通过基于信号序列优化机器听觉模型的研究,为蜂群健康与活动状态的监测提供依据。【方法】在蜂箱内设置音频传感器,以非侵入性和无干扰性的方式持续记录6类蜂群音频,针对传统的音频分类方法中未考虑时序信息和分类准确度不高等问题,提出一种基于双向长短期记忆(bidirectional long short-term memory, BiLSTM)网络优化的多分类模型。基于梅尔频率倒谱系数提取音频特征,并构建以BiLSTM为基准的蜂群状态分类模型;引入卷积神经网络(convolutional neural network, CNN)和自注意力机制(self-attention mechanism, SA)对BiLSTM的输入和输出进行优化;构建优化的CNN-BiLSTM-SA模型用于6类蜂群状态的精准识别。【结果】与CNN和BiLSTM模型相比,CNN-BiLSTM-SA模型的分类准确率最高,训练集和验证集准确率均大于0.990 0,测试集准确率为0.988 6,交叉验证平均准确率为0.981 5。【结论】CNN-BiLSTM-SA模型为蜂箱内蜂群状态精准识别提供了有效技术支持,有助于未来智能养蜂和音频传感监控的发展。【Objective】The study aimed to provide a basis for monitoring the health and activity status of honeybee colonies by opti⁃mizing machine auditory models based on signal sequences.【Method】Audio sensors were set up to acquire the auditory information of 6 types of bee colony in a non⁃invasive and interference⁃free way.A multi⁃classification model based on the optimization of bidi⁃rectional long short⁃term memory(BiLSTM)was proposed to address limitations such as overlook of temporal information and lac⁃king accuracy in traditional audio classification methods.Audio features were extracted based on Mel⁃frequency cepstrum coeffi⁃cients,and then were applied to establish the bee colony status classification models based on BiLSTM.Convolutional neural network(CNN)and self⁃attention mechanism(SA)were introduced to optimize the inputs and outputs of BiLSTM to construct the optimized CNN⁃BiLSTM⁃SA model for the accurate detection of status of 6 types of bee colony.【Result】Compared with CNN and BiLSTM models,the proposed CNN⁃BiLSTM⁃SA model resulted in the highest classification accuracy,with the accuracy of training and validation sets greater than 0.990 0, accuracy of test set at 0.988 6, and average accuracy of 0.981 5 in cross⁃validation. 【Conclusion】CNN⁃BiLSTM⁃SA model plays an important role in the accurate recognition of bee colony status in beehives, which is contributed tothe further development of intelligent beekeeping and audio sensing monitoring.

关 键 词：蜂群状态 机器听觉 双向长短期记忆 卷积神经网络 自注意力机制 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]