An Exploratory Study of Low Cost PM2.5 Sensor Calibration Method

[[abstract]]如何維持好的空氣品質是現今社會熱切關注的議題。為了有效地勸導民眾減少排放汙染物的行為，廣布空氣品質微型監控設備來快速找出汙染源是許多環保團體努力的目標。隨著微型監控設備的增加，對低成本 PM2.5感測器的數據品質，各界也產生越來越多的疑慮，因此本研究將重心放在感測器校正問題的探討。為了建立校正效果最好的機器學習模型，本研究使用了六種資料清理方法和四種參數最佳化演算法，以及本研究提出的三種參數挑選方法與兩種模型訓練方法，並且能在最後得到與環保署自動監測站有非常接近的數據品質。然而實驗結果顯示並不存在一個方法，能夠訓練出適合所有微型監控設備的校正模型，所以本研究認為不同微型監控設備都應該採用自動最佳化排程來嘗試各種方法的組合，並從中找出最合適的校正模型。[[abstract]]Nowadays, people concerned about the issue of how to maintain clean air. For persuading people to reduce pollutant emission, lots of environment protection teams devoted to find pollution source quickly by deploying micro monitoring devices widely. With the growth of micro monitoring device, more and more doubt arising at quality of low cost PM2.5 sensor data from community. As the result, this study focus on exploring problem of sensor calibration. In order to build machine learning model with the best result of calibration, this study using six methods of data cleaning and four methods of hyperparameter optimization algorithm. In addition, this study proposes three methods of hyperparameter selection and two methods of model training to build the best model. Eventually, the quality of calibrated data can be close to EPA’s automatic monitoring station. However, experiment result show that there is not exist a method which can train the model calibrating well on different micro monitoring device. Therefore, this study think that it should try different combination of methods automatically via an automatic optimization pipeline to find the best calibration model on different micro monitoring device.[[note]]碩