台灣近年來 PM2.5 空氣汙染的議題逐漸受到重視，增設了許多價格 較為低廉的感測器，但是這些感測器容易受到環境因素影響造成較大的 誤差，加上數量龐大造成每台感測器的維護頻率低，單一區域感測器回 傳的數值不如國家級測站來得可靠， 本論文比較了監督式、無監督式、及半監督式的演算法在偵測異常 傳感器的效果。為了結合感測器的時空資訊，我們將監測值轉成圖片資 料、整合性資料、以及整合資料結合時序資料來準備訓練數據。我們根 據工業技術研究所提供的檢測記錄得到感器測的狀態值（正常或異常）， 探討了標記資料的比例對半監督模型預測效能的影響。實驗結果顯示： 我們研究的方法優於目前的隨機巡檢機制。;The PM2.5 issue has drawn much attention in Taiwan, and many inexpensive sensors have been deployed in recent years. However, these sensors are fragile and susceptible to environmental factors. In addition, the large number of sensors results in low maintenance frequency, so the monitored values returned by a single sensor are unreliable. This thesis compares supervised, unsupervised, and semi-supervised methods to identify the problematic sensors. We prepared the training data by converting monitored values into images, integrated data, and se quential data to incorporate the spatio-temporal information of the sensors. We obtained sensors’status (normal or abnormal) based on the inspec tion records provided by the Industrial Technology Research Institute. We explored how the ratio of labeled data to unlabeled data influences the per formance of the semi-supervised models. Experimental results show that our studied methods outperform the current inspection strategy (random inspection).
