基于采食行为参数的奶牛个体采食量量化方法研究  

Quantification method of dairy cow individual feed intake based on dairy cow feeding behavior parameters

作　　者：魏晓莉[1] 刘慧鑫 刘慧环 寇胜利 李晓 沈维政[1] WEI Xiaoli;LIU Huixin;LIU Huihuan;KOU Shengli;LI Xiao;SHEN Weizheng(School of Electrical Engineering and Information,Northeast Agricultural University,Harbin 150030,China;Lindian County Animal Husbandry and Veterinary Bureau,Daqing Heilongjiang 166300,China;Heilongjiang Dairy Industry Association,Harbin 150069,China;China Agriculture Press,Beijing 100125,China)

机构地区：[1]东北农业大学电气与信息学院,哈尔滨150030 [2]黑龙江省林甸县畜牧兽医局,黑龙江大庆166300 [3]黑龙江省奶业协会,哈尔滨150069 [4]中国农业出版社,北京100125

出　　处：《东北农业大学学报》2023年第6期42-52,共11页Journal of Northeast Agricultural University

基　　金：国家重点研发计划项目(2022YFD1301104);黑龙江省博士后科研启动基金项目(LBH-Q21062);财政部和农业农村部:国家现代农业产业技术体系资助(CARS36)。

摘　　要：受饲养环境和设备限制,现存奶牛个体采食量监测方法难以应用于商业生产。为找到简单高效的个体采食量量化方法,以采食行为参数(采食时间、躺卧时间、行走步数)、个体体重及日粮粗精比作为输入参数,构建4种采食量量化模型:反向传播神经网络(BP)模型、遗传算法优化BP神经网络(GA-BP)模型、粒子群算法优化BP神经网络(PSO-BP)模型、多项式学习率衰减算法优化BP神经网络(PDLR-BP)模型,并经模型评价验证找到适宜的算法函数。结果表明,结合采食行为数据与改进型BP算法可实现奶牛个体采食量量化研究,4个模型建模精度高,模型量化值与实测值呈显著相关,决定系数高于0.90;其中,PSO-BP模型精准性最优且收敛更快,模型MSPE,MAE和R2值分别为0.046 kg^(2)·d^(-1)、0.166 kg·d^(-1)和0.95,其均方相对误差RMSE=0.214 kg·d^(-1)分别低于BP模型、GA-BP模型和PDLR-BP模型10%、9.3%和6.9%。PSO-BP模型极值差降幅明显,说明PSO-BP模型稳定性更优。Existing methods for monitoring individual feed intake are limited by the feeding environment and equipment,making it difficult to apply them to commercial production.In order to find a simple and efficient method for quantifying the feed intake of dairy cows,the parameters of feeding behavior(eating time,lying time and walking steps),individual body weight and dietary ratio of concentrate to roughage were used as input parameters to establish four feed intake prediction models:BP(back propagation)artificial neural network,GA-BP(genetic algorithm-optimized back propagation)artificial neural network,PSO-BP(particle swarm-optimization back propagation)artificial neural network and PDLR-BP(polynomial decay learning rate-optimization back propagation)artificial neural network,and find the appropriate algorithm function through model evaluation and verification.The results showed that the combination of individual feeding behavior data and the improved BP algorithm could achieve the quantification of individual feed intake of cows,and four models had high accuracy,and the quantified values of the models were significantly correlated with the measured values,with the R2(coefficients of determination)above 0.90;among them,the PSO-BP model had the best accuracy and faster convergence,with model MSPE(mean square percentage error),MAE(mean absolute error)and R^(2) values of 0.046 kg^(2)·d^(-1),0.166 kg·d^(-1) and 0.95,respectively,and its RMSE(mean square relative error)=0.214 kg·d^(-1) compared with BP network,GA-BP and PDLR-BP models was reduced by 10%,9.3%and 6.9%,respectively.At the same time,the relative error change trend chart showed that the extreme value difference of the PSO-BP model decreased significantly,which further proved that the PSO-BP model had better stability.

关 键 词：奶牛 可穿戴设备 行为参数 算法优化 个体采食量量化 

分 类 号：S823[农业科学—畜牧学]