利用機器學習在物聯網系統之感知器異常偵測與修復

Abstract

物聯網蓬勃的發展，種類繁多的傳感器可以搜集各種類的資訊，如溫度、濕度、震 動等等的訊息。而這些訊息可以用在各種不同的場合，設定不同的條件可以執行不同的 動作、回應，甚至可以傳輸至雲端計算中心，進而做大數據分析，萃取有用的資訊。隨 著感知器數量飛速的成長，目前也有越來越多大型的感知器系統出現，像是大規模的氣 候偵測、智慧電網、工廠的機器檢測等等的應用。而這也代表著傳感器會被擺放在各種 不同的環境，傳感器的可靠度可能因外在環境或是本身老舊程度有所變化。可能會搜集 到錯誤的資訊，或是資料的遺漏，導致錯誤的判斷，因此偵測傳感器的健康狀況也成為 了非常重要的議題。 在本論文中，提出了基於機器學習，偵測傳感器健康狀況的系統架構。藉由附近不 同傳感器的資訊，可以推測出目標傳感器應有的值。在傳感器發生異常時馬上知道，甚 至快壞掉之前發出警告。此外，我們分析了有異常之傳感器的歷史資料，能夠歸納出此 傳感器錯誤的種類，進而使用軟體的方式加以修正。使得傳感器系統的可靠度更加提 升。

With the growth of the Internet of Things (IoT), the number of IoT sensors grows fast, and it will produce a great amount of data, which can be analyzed then used by governments or businesses. And there are a lot of sensing systems in various environments, such as environmental monitoring, machine-condition monitoring, smart power grid systems, etc, but sensors could be unreliable due to either improper calibration, low battery levels or hardware failures of the device. Unreliability may cause users to make erroneous decisions or inaccurate analysis. Therefore, it is important to detect if sensor is abnormal as earlier as possible. In this thesis, we propose a system architecture that can detect the abnormality among the sensors, based on machine learning. This architecture can detect the faulty sensors in real-time, by exploring the correlation among the sensors, and predict the expected value via other correlated sensors. Besides, we analyze the faulty sensor’s data pattern, in order to classify the fault type of faulty sensors. Furthermore, we try to recover the faulty sensors by software, improves the reliability of sensing system.