整合多重感測訊號於蜂巢健康狀態之監測與預測

Abstract

蜜蜂是重要的昆蟲授粉媒介，除協助提高經濟作物產量外，同時在維持生態系穩定上扮演重要角色。監測蜜蜂行為和蜂箱健康狀況不僅對了解蜜蜂生物學至關重要，也有利於協助養蜂人進行蜂箱健康管理。基於以上理由，我們建立智慧蜂箱監測系統，目標提升蜂箱管理品質並降低蜂箱損失的風險。我們使用多感測器組建智慧蜂箱健康狀態監測系統，以監測多項蜂箱特徵，包括巢內溫濕度、重量、蜜蜂交通量和聲音信號。本研究收集了南投水里與雲林古坑兩地共四蜂箱的長期數據集，自多重感測器與當地氣象站所蒐集的資料中，導出27項蜂巢特徵，以研究蜂巢特徵用於健康狀態偵測的可能性。重量特徵方面，我們自每日重量變化曲線中，以分段迴歸之方式提取蜜蜂白日出巢與夜間歸巢之時間點以觀察蜜蜂生理活動週期，並對照蜜蜂活動量特徵與聲音特徵之日夜週期，驗證各觀測值之間的關連與正確性。聲音訊號方面，我們使用聲景指數 (soundscape indices) 中的聲音複雜度指數 (Acoustic complexity index, ACI)、聲音多樣性指數 (Acoustic diversity index, ADI)、聲音均勻度指數 (Acoustic evenness index, AEI)、頻譜熵 (Spectral entropy)、與方均根能量 (RMS Energy) 等五種來量化巢內錄音訊號，並細部檢視一日內各指數的變化情形。我們善用智慧蜂箱多角度觀測的優勢，結合環境溫濕度、雨量、巢重變化與蜜蜂進出量資訊，綜合檢視在降雨、低溫、收穫等三種狀態下聲景指數的變化，細部且直觀的分析環境變化、晝夜變化與蜂巢蜂鳴改變間的連動關係。基於對以上蜂巢特徵的了解，我們以隨機森林模型 (Random Forest) 對資料集進行訓練，得聲音特徵群為二分類蜂巢健康狀態預測命題下之最佳分類特徵群，準確度達0.80。若以各巢單獨訓練模型，準確度能進一步達0.84。多重感測智慧蜂箱監控系統能自動收集長期數據，監測蜂箱狀態，幫助了解蜜蜂生理活動與長期變化規律，並能提出蜂巢衰敗警示，進一步幫助養蜂人以數據導向的方式管理蜂箱，對養蜂業之智慧化做出推動與貢獻，協助蜂農良好管理蜂巢崩解之風險。

Honeybees are important insect pollinators ensuring food security and maintaining the biodiversity of ecosystems. Monitoring honeybee behavior is not only essential for understanding the biology of the honeybee but also beneficial for beekeepers for beehive management. In the apiculture industry, beekeepers usually look after beehives regularly and manually. However, the assessment of beehive health status is laborious and requires considerable experience. An automated beehive monitoring system will facilitate efficient beehive management and reduce the risk of beehive losses. We propose an intelligent beehive health status monitoring system using multiple sensors and a sensor fusion technique. The system monitors various features of beehives, including in-hive temperature, humidity, hive weight, bee traffic, and acoustic signals. A long-term dataset of 4 beehives in two different locations was collected. 27 features were derived from sensors’ data to aid in hive health status detection and prediction. We analyzed the relationship between weight features and beehive circadian rhythm, and correlated them with bee traffic features and acoustic features to verify their correlation and correctness. Harvest status was also explored by assessing the weight change in the daily weight curve. Regarding the aspect of acoustic features, soundscape indices are used to interpret the acoustic signals, including acoustic complexity index (ACI), acoustic diversity index (ADI), acoustic evenness index (AEI), spectral entropy, and root mean square energy (RMS energy). We inspected these indices under raining, cold, and harvest situations. Leveraging the multi-sensing dataset, we demonstrated a clear relation between environmental changes and beehive acoustic changes. Based on these features, the two-class hive health status prediction with Random Forest model hit an accuracy of 0.80. The multi-sensor intelligent beehive monitoring system is an efficient tool to observe the long-term hive activity and to help beekeepers in managing their beehives in a data-driven approach thereby improving the beekeeping quality.