关键词: 猪只行为, 部署位置, 可穿戴设备, 姿态传感器, Relief算法, 卷积神经网络

Abstract: Pig behavior monitoring is a key technology to realize intelligent and refined pig breeding. In this study, two wearable devices were designed by using posture sensors, which were deployed on the backs and necks of experimental pigs to collect data on seven daily behaviors, including lying prostrate, side lying, feeding, standing, walking, dog sitting and excretion. There are 3 types of behavioral datasets collected, which are divided into single‑source deployment (back or neck) and combined deployment (back+neck). The collected signals were processed by wavelet domain denoising, data signal selection, data segmentation, time‑domain feature extraction, and feature selection of Relief algorithm, in which the feature selection of Relief algorithm combined with BP neural network and random forest algorithm to determine the feature retention number k, and then the processed data were sent to four classification models of BP neural network, random forest, convolutional neural network and extreme learning machine for analysis. The overall accuracy of the combined deployment (back+neck) was significantly higher than that of single‑source deployment (back or neck). The classification accuracy of behavior by using BPNN, RF, CNN, and ELM in the combined deployment was 90.59%, 87.14%, 91.67%, and 82.5%. In this study, the characteristics of the duration of each behavior in pigs were analyzed, and the classification effect of different behaviors required different evaluation indexes for analysis. Combining various evaluation indicators, the classification model deployed by combination is the best CNN convolutional neural network. The combined deployment and fusion of the two attitude sensors can classify pig behavior, and the results of this study are of great significance for the daily behavior monitoring and health assessment of pigs.

Key words: pig behavior, deployment location, wearable devices, posture sensors, Relief algorithm, CNN