应烨伟, 曾松伟, 赵阿勇, 颜菲菲. 基于颈环采集节点的母羊产前行为识别方法[J]. 农业工程学报, 2020, 36(21): 210-219. DOI: 10.11975/j.issn.1002-6819.2020.21.025
    引用本文: 应烨伟, 曾松伟, 赵阿勇, 颜菲菲. 基于颈环采集节点的母羊产前行为识别方法[J]. 农业工程学报, 2020, 36(21): 210-219. DOI: 10.11975/j.issn.1002-6819.2020.21.025
    Ying Yewei, Zeng Songwei, Zhao Ayong, Yan Feifei. Recognition method for prenatal behavior of ewes based on the acquisition nodes of the collar[J]. Transactions of the Chinese Society of Agricultural Engineering (Transactions of the CSAE), 2020, 36(21): 210-219. DOI: 10.11975/j.issn.1002-6819.2020.21.025
    Citation: Ying Yewei, Zeng Songwei, Zhao Ayong, Yan Feifei. Recognition method for prenatal behavior of ewes based on the acquisition nodes of the collar[J]. Transactions of the Chinese Society of Agricultural Engineering (Transactions of the CSAE), 2020, 36(21): 210-219. DOI: 10.11975/j.issn.1002-6819.2020.21.025

    基于颈环采集节点的母羊产前行为识别方法

    Recognition method for prenatal behavior of ewes based on the acquisition nodes of the collar

    • 摘要: 针对目前母羊产前行为监测费时费力、精准度较低、可识别行为类型单一等问题,以颈环采集节点获得的加速度数据为研究对象,提出了一种基于区间阈值与遗传算法优化支持向量机(Genetic Algorithm-Support Vector Machine, GA-SVM)分类模型的母羊产前行为识别方法。该方法首先对合加速度数据进行小波降噪和提取轮廓线预处理,再利用区间阈值分类法,识别出行走、趴卧2种行为。在此基础上,采用GA-SVM分类模型实现饮水、采食、反刍3种行为的准确识别。试验表明,颈环采集节点能够实时采集和传输产前母羊颈部的活动信息,并能有效区分5种产前行为活动。可识别行为类型增多,且在适用性方面有了较大提升,平均准确率达到97.88%,比传统的决策树算法、K最近邻(K-Nearest Neighbor, KNN)算法、支持向量机(Support Vector Machine, SVM)算法分别提高了31.26%、21.87%、21.9%。该研究对建立产前母羊运动量及健康评估模型、提高繁殖及工作效率、实现智能化管理等方面具有十分重要的意义。

       

      Abstract: Aiming at the current time-consuming and manpower-spending problems of prenatal behavior monitoring of ewes, a monitoring system based on the acquisition node of the collar was designed. The acquisition node of the self-made collar was integrated with the MPU6050 sensor, which collected prenatal behavioral acceleration information of ewes in real-time. The collected acceleration data was wirelessly transmitted to the embedded base station through Zigbee technology. The server received the data from the GPRS module in the base station, and then the data was stored in the MySQL database. Finally, the data was displayed on the webpage or mobile phones. According to the monitoring video of ewes, the collected acceleration data was calibrated at the same time, which provided sufficient samples for the behavior classification models. Besides, aiming at the problems of low accuracy and less recognizable behaviors of ewes in labor, regarding the collected acceleration data as the research object, a recognition method for prenatal behavior of ewes was proposed based on the classification model of the interval threshold and the Genetic Algorithm-Support Vector Machine (GA-SVM). In this method, three-dimensional acceleration data was firstly synthesized into one-dimensional resultant acceleration data, and then the resultant acceleration data was preprocessed by noise reduction of db5 wavelet and extraction contour of the sliding window. According to its amplitude fluctuation characteristics, the interval threshold classification method was used to recognize the two behaviors of walking and lying. In the light of the obvious characteristics of the spatial clustering effect for the three-axis acceleration data of drinking, eating, and ruminating behavior, the Genetic Algorithm (GA) was used to search the global optimal solution of penalty parameter and kernel function parameter of the Support Vector Machine (SVM) to accurately recognize the three behaviors. Finally, through calculating the four indicators including accuracy, precision, sensitivity, F1-score based on the confusion matrix, and then the performance between the classification method proposed in the research and several common classification algorithms were evaluated. The comparison results illustrated that the F1-score of the classification method proposed in the research was inversely proportional to the movement amplitude of ewes, and its precision was also inversely proportional to the sample size. The average accuracy of the classification method proposed in this work was 97.88%, which was 31.26%, 21.87%, and 21.9% higher than the traditional decision tree algorithm, the K-Nearest Neighbor (KNN) algorithm, and the SVM algorithm, respectively. The main reason was that with the increase of behavior recognition types, using only one feature vector and classification algorithm, it was difficult to find the most suitable partition hyperplane, resulting in the unavoidable decline in the correct recognition rate of certain behaviors. The experimental results showed that the monitoring system for the prenatal behavior of ewes based on the acquisition node of the collar could collect and transmit the activity information of prenatal ewes’ neck simultaneously. Besides, the classification method based on the interval threshold and GA-SVM could effectively distinguish five kinds of prenatal activities including drinking, eating, ruminating, walking, and lying. The number of identifiable behavioral types increased, and the applicability was greatly improved. This work also set up a control group to conduct a comparative experiment and found that the prenatal ewes had no abnormal behaviors after the 24 h adaptive period. Therefore, the acquisition node of the self-made collar would not affect the prenatal behaviors of ewes. The results were of great significance for establishing the amount of exercise and health assessment models of prenatal ewes, improving the efficiency of reproduction, and realizing intelligent management.

       

    /

    返回文章
    返回