基於多元感測與機器學習之乳牛泌乳相關健康與行為監測系統之開發

Abstract

本研究的目標是開發一套乳牛健康監測系統，架設多項感測器並利用資料融合與機器學習技術，以實現智慧農業管理的目的。利用小樣本學習演算法於深度學習模型，建立牛臉特徵向量提取模型，監測個體牛隻採食時長。同時，結合牛臉辨識演算法和熱成像監測系統，建立個體乳牛眼睛溫度監測工作模組，並使用輕量化模型進行邊緣運算。開發模型更新自動化演算法，以解決牛隻頻繁更換和淘汰的問題，也降低了訓練模型所需的人力及時間成本。牛眼溫度監測系統採用串流時機選擇演算法，實現自動化監測牛眼溫度，將儲存空間的佔用量減少原本的80百分比。此外，本研究建立了Docker環境，建置多項服務，以收集和整合感測器資訊，架設工作模組之間的通訊管道，使整個系統可以快速重建、管理及更新。在本研究中，整合了計算採食時長的影像監測系統、測量牛眼溫度的熱成像工作模組、監測乳牛呼吸頻率的雷達系統、收集運動行為特徵的慣性測量單元，以及測量環境因素的溫濕度感測器。利用這些感測器收集乳牛在生理現象、運動行為和活動環境等方面的資訊。透過將多重感測器的資訊進行整合，本研究建立乳牛綜合健康指標和熱緊迫診斷平台。這套系統能夠根據不同的指標，對乳牛的健康狀況進行監測和評估。利用移動平均方法減少每日採食時長的雜訊，解決乳牛個體間的差異性，分析其採食狀況是否異常。後續進行多項相關性分析，發現個體牛隻的採食時長與溫濕度指數呈負相關，整體相關係數為0.37，群體牛隻的呼吸頻率則與溫濕度指數呈正相關，相關係數為0.57。除此之外，我們利用非監督式機器學習模型將多重感測資訊進行分類，得到最佳轉折點於THI值70.7及78.0，進一步診斷乳牛的熱緊迫現象的嚴重程度。我們開發了乳牛發情期偵測演算法，建立相關規則並進行實驗優化參數，其F1-score為0.833。最後，將所有資訊呈現於使用者介面上，提供牧場人員快速判讀牧場乳牛的健康狀況。

The objective of this study was to develop a dairy cow health information system by deploying multiple sensors and utilizing data fusion and machine learning techniques. The system aimed to establish comprehensive health indicators and a heat stress diagnostic platform for dairy cows, with the purpose of achieving smart agricultural management. A deep learning model was developed using a small sample learning algorithm to extract facial features of cows, allowing for monitoring of individual cow's feeding duration. Additionally, a cow facial recognition algorithm and thermal imaging monitoring system were combined to create an individual cow eye temperature monitoring module, employing lightweight models for edge computing. An automated model updating algorithm was developed to address the frequent replacement and culling of cows, reducing the human and time costs required for model training. The cow eye temperature monitoring system adopted a streaming time selection algorithm, enabling automated monitoring of cow eye temperature and reducing storage space consumption from 1.3GB per day to 260MB. Furthermore, a Docker environment was established in this study to build multiple services for collecting and integrating sensor information. It also facilitated the communication pipeline between different modules, allowing for quick reconstruction, management, and updates of the entire system. The study integrated an image monitoring system for measuring feeding duration, a thermal imaging module for measuring cow eye temperature, a radar system for monitoring cow respiratory rate, an inertial measurement unit for collecting motion behavior characteristics, and temperature and humidity sensors for measuring environmental factors. These sensors were used to collect information on physiological phenomena, motion behavior, and environmental conditions of dairy cows. By integrating the information from multiple sensors, this study established comprehensive health indicators and a heat stress diagnostic platform for dairy cows. The system enabled monitoring and evaluation of the health condition of dairy cows based on different indicators. The use of moving average methods reduced noise and addressed inter-individual variability to analyze whether the feeding behavior of dairy cows was abnormal and establish corresponding indicators. Subsequent correlation analysis revealed a negative correlation coefficient of 0.61 between individual cow feeding duration and the temperature-humidity index (THI), while the respiratory rate of the entire herd showed a positive correlation coefficient of 0.75 with the THI. Furthermore, an unsupervised machine learning model was employed to classify the information from multiple sensors, determining the optimal turning points at THI values of 70.7 and 78.0. This further allowed for the assessment of the severity of heat stress in dairy cows. A heat detection algorithm for dairy cow estrus period was developed, with related rules established and experimental parameter optimization conducted, resulting in an F1-score of 0.833. Finally, all the information was presented on a user interface using Grafana, enabling dairy farm personnel to quickly review the health status of cows on-site.