基於多元感測與機器學習之智慧蜂箱健康狀態預測系統

Abstract

本研究開發一套基於智慧蜂箱的蜂群健康監測系統，利用多感測器長期蒐集蜂箱內外部環境資料，結合時間序列預測與異常分類模型，實現蜂群重量、進出流量及花粉採集率的精確預測與健康狀態異常偵測。資料來源為 2024 年 9 月至 2025 年 4 月於台灣雲林兩處場域蒐集的感測數據與蜂群檢查表，蜂農每週填寫健康評估表單以標註蜂群狀態，作為機器學習模型的訓練與驗證依據。模型部分採用 GRU 與 SARIMAX 進行單變量與多變量時間序列預測。結果顯示多變量模型在三項目標變數的預測精度均優於單變量模型，其中 GRU 模型在跨蜂箱驗證下的誤差表現最佳，蜂群重量的平均絕對百分比誤差（MAPE）為 0.6%～2.1%，進出流量的 MAPE 為 19%～24.6%，花粉採集率的對稱平均絕對百分比誤差（sMAPE）為 16.7%～31.3%。透過 SHAP 分析進一步辨識影響預測結果的關鍵特徵，模型未出現預測停滯或漂移現象，具備長期穩定性。GRU 二元分類模型在蜂群健康狀態預測的整體準確率達 94%，能有效偵測失王、蜂巢壅擠等關鍵異常事件。系統能夠提供穩定且可靠的數據化蜂群動態洞察，支援早期健康問題識別與精準養蜂管理，提升蜂群存活率與管理效率。

This study presents a smart beehive-based health monitoring system that combines multi-sensor data collection, time-series forecasting, and anomaly detection to monitor honeybee colony health. Data were collected from September 2024 to April 2025 at two apiaries in Yunlin County, Taiwan. Sensor measurements and beekeeper-completed evaluation forms were used to label colony health status for training and validating machine learning models. Two forecasting models, GRU and SARIMAX, were developed to predict hive weight, bee traffic, and pollen collection rate under both univariate and multivariate settings. Experimental results showed that multivariate models consistently outperformed univariate ones, with the multivariate GRU achieving the most stable and generalizable performance. Specifically, during cross-hive validation, the forecasting errors were: a Mean Absolute Percentage Error (MAPE) of 0.6%–2.1% for weight prediction, 19%–24.6% MAPE for bee ingress–egress traffic, and a Symmetric Mean Absolute Percentage Error (sMAPE) of 16.7%–31.3% for pollen collection rate. SHAP-based interpretability analysis further identified the most influential features contributing to model accuracy, and the model exhibited no signs of prediction stagnation or shifting. Additionally, a GRU-based binary classifier for colony health status achieved an overall accuracy of 94%, effectively detecting critical anomalies such as queen loss and hive overcrowding. Overall, the proposed system reliably provides data-driven insights into colony dynamics, enabling early detection of health issues and supporting precision apiculture practices for improved hive management.