特征提取和特征降维在猪只行为分类识别中的应用  

Application of feature extraction and feature dimension reduction in pig behavior identification and classification

作　　者：靳敏[1] 杨博文 王春光[3] JIN Min;YANG Bowen;WANG Chunguang(College of Electrical Engineering,Changzhou Institute of Mechatronic and Electrical Technology,Changzhou 213164,China;College of Mechanical Engineering,Changzhou Institute of Mechatronic and Electrical Technology,Changzhou 213164,China;College of Mechanical and Electrical Engineering,Inner Mongolia Agricultural University,Hohhot O10018,China)

机构地区：[1]常州机电职业技术学院电气工程学院,江苏常州213164 [2]常州机电职业技术学院机械工程学院,江苏常州213164 [3]内蒙古农业大学机电工程学院,呼和浩特010018

出　　处：《黑龙江畜牧兽医》2022年第23期53-59,134,共8页Heilongjiang Animal Science And veterinary Medicine

基　　金：“十二五”国家科技支撑计划项目(2014BAD08B05);内蒙古自治区研究生科研创新项目(B2018111948)。

摘　　要：为了有效改善猪只行为分类识别效果,试验采用三轴加速度传感器获取试验猪(猪A、猪B、猪C)在X轴、Y轴和Z轴三个方向上的加速度数据,建立试验猪只行为数据集,分别提取X轴、Y轴和Z轴的平均值、中位数、最大值、最小值、第一四分位数和第三四分位数,共同构成一个包含21个特征在内的数据集,分别采用ReliefF算法和随机森林算法就各特征对试验猪行为分类识别结果影响的大小进行分析与排序,删除与分类识别性能相关性小的特征,将21维数据集降维至9维。结果表明:将经ReliefF算法降维的数据集用于猪只行为识别与分类,猪A、猪B猪C的总体平均准确率分别为80.9%、81.7%和82.0%;将经随机森林算法降维后的数据集用于猪只行为识别与分类得到的总体平均准确率分别为86.4%、85.3%和87.2%。说明采用随机森林算法进行特征降维的效果更好,更适用于处理猪只行为数据。To improve the identification and classification effect of pigs’ behavior, a triaxial acceleration sensor was used to obtain the acceleration data of experimental pigs(A, B and C) in the X-axis, Y-axis and Z-axis to establish the behavioural data set of the experimental pigs. The mean, median, maximum, minimum, the first quartile and the third quartile of X-axis, Y-axis and Z-axis were extracted respectively to form a data set containing 21 features. ReliefF algorithm and Random Forest algorithm were used to analyse and rank according to the influence of each feature on the identification and classification results of experimental pig behavior. Then, the features with little correlation in identification and classification performance were deleted, and the dimension of the 21-dimensional data set was reduced to 9 dimensions. The results showed that the overall average accuracy of pigs A, B and C was 80.9%, 81.7% and 82.0%, respectively, when the dimensionality reduction dataset obtained by ReliefF algorithm was used for pig behavior identification and classification. Similarly, the overall average accuracy of pig behavior identification and classification was 86.4%, 85.3% and 87.2%, respectively. The results indicated that the effect of feature dimension reduction obtained by the Random Forest algorithm was better, and was more suitable for processing the behavior data of pigs in this study.

关 键 词：三轴加速度传感器 特征提取 特征降维 分类识别 人工神经网络 

分 类 号：S828[农业科学—畜牧学]