強化農業生產環境安全的水質監控機制

Abstract

為了解決食安之問題，相關單位透過管理平台回收現地水質連續自動監測數據，加強對農業生產環境之掌握，然而現行採用的現場自動化重金屬測站檢測成本昂貴，除了無法廣設監測站，頻繁地進行採樣亦有其困難之處，造成測站所獲得的訊息可能會有空窗期。 本論文之目標為利用有限的資訊建立一套預警系統，使用機器學習以及深度學習之技術，針對水中重金屬濃度的一系列分析，來預測其灌溉水可能被污染的時間點，以能夠及早進行任何有利環境之行動。預警系統分成以下部分，首先是預測重金屬含量是否超過標準，利用來自於測站的水質資料，使用機器學習模型進行灌溉水質的超標預測；然而上述辦法卻無法有效掌握水質濃度隨著時間之變化，於是加入了深度學習時間序列的分析，來為重金屬濃度的趨勢預測；天氣狀況亦也是會影響工廠偷排廢棄物的因素之一，因此於系統中加入氣象預測資訊，包含降雨機率、天氣描述。最後綜合以上之特性，利用預測結果得出一個預警分數，監控平台的管理者可依照其分數來決定其接下來的行動，而經過一系列的驗證，可發現上述模型皆獲得有效之預測。

To solve the problem of food safety, related departments have taken some actions to enhance the management of the production environment. They collect water quality data through sampling stations. However, the expense of the heavy metal sampling is costly and thus unable to construct a wide range of heavy metal sampling stations. Besides, they also can not collect data frequently. Therefore, the main purpose of the thesis is to solve the above problem by establishing a supervision mechanism. The mechanism predicts whether the water will be polluted by machine learning, and then informs the supervisor to take actions. First, the mechanism predicts the excess of heavy metal concentration. This part is trained by the data of the water quality stations through machine learning techniques. But it is hard to know the difference of the concentration between different time. The time series characteristic of RNN has added to the solve the disadvantage to estimate the concentration. In addition to the above part, we also take the weather forecast into consideration. Finally, calculating a score after making predictions, and getting a comprehensive performance through combining the above parts. The result shows that the models have good performances.