人工智慧應用於消費級氣體感測器陣列之校正

Abstract

近年來空氣污染逐漸嚴重，空污議題逐漸受到重視，而為了監控有害氣體濃度，氣體感測器布置密度也必須提高。高單價感測器雖然精準，但大規模設置，成本會過高。消費級氣體感測器的優點在於設置成本低廉，但其性能則必須有進一步的加強。消費級氣體感測器具有對溫濕度產生變化、對多種氣體有反應、使用時的重現性不佳及感測器個體差異等問題。解決性能上問題後，消費級氣體感測器的實用性可大幅提升，可以提高量測氣體濃度的準確性。本研究將量測SO₂與CO的濃度，並使用人工智慧校正消費級氣體感測器，使感測器表現優於原廠提供的校正，以增加可用性。深度學習為人工智慧的一部分，具有自動提取資料特徵並解決問題的特點。本研究將使用人工智慧進行氣體感測器校正。 首先分別建立SO₂感測器校正模型以及CO感測器校正模型。在密閉腔室通入SO₂或是CO氣體搭配不同溫度濕度量測，並記錄感測器響應建立人工智慧數據庫，作為兩種氣體感測器校正模型所需的訓練及測試資料。使用原廠校正時，SO₂平均誤差2.58 ppm、CO平均誤差2.06 ppm。實驗顯示單一SO₂感測器校正模型平均測試誤差1.5 ppb，單一CO感測器校正模型平均測試誤差19.8 ppb，顯然優於原廠校正。接著建立氣體感測器陣列校正模型，在密閉腔室通入SO₂和CO的混合氣體搭配不同溫度濕度，並記錄感測器響應建立人工智慧數據庫，作為訓練及測試資料。使用深度類神網路進行校正並找出較好的模型結構以及資料擷取時間，此模型SO₂平均測試誤差8 ppb、CO平均測試誤差166 ppb。 接著使用類神經網路隱藏層的模組化以及輸入資料kernel trick非線性增維的方法建立更強的模型，強化後的模型SO₂平均測試誤差降至1.7 ppb、CO平均測試誤差降至77 ppb。量測數據使用原廠所提供的校正時，SO₂濃度平均誤差達2.65 ppm、CO濃度平均誤差達2.69 ppm，使用感測器陣列校正模型明顯降低誤差。最後是氣體感測器個體差異的校正，首先建立通用型的SO₂感測器校正模型，接著建立可替換SO₂感測器之氣體感測器陣列模型。將四個不同的SO₂感測器量測數據正規化後進行模型訓練，並使用一個從未訓練過的SO₂感測器數據進行測試，其測試誤差為4.35 ppb。將通用型的SO₂感測器校正模型與氣體感測器陣列校正模型結合而成可替換SO₂感測器之氣體感測器陣列模型，接著使用新的SO₂感測器所量測的混合氣體數據測試，SO₂平均測試誤差109 ppb、CO平均測試誤差2450 ppb，對比原廠校正成功降低了誤差，且此模型可以替換SO₂感測器，其通用性對廣布氣體感測器陣列的需求來說相當有幫助。 總體而言，人工智慧可以提升消費級氣體感測器之精準度，消費級氣體感測器可以在環境中進行有害氣體偵測並達到監控空氣品質之目的。

In order to monitor the concentration of harmful gases, the layout density of gas sensors must be increased. Although the high precision gas sensor is accurate, the setup cost will be too high, so a cheaper option is needed. The advantage of the consumer-grade gas sensor is that the setup cost is low, but its performance needs to be enhanced. Consumer-grade gas sensors have problems such as value disturbed by temperature and humidity, response to a variety of gases, poor reproducibility, and individual sensor differences. After solving the performance problem, the practicability of the consumer-grade gas sensor will be significantly improved, and it can accurately measure the gas concentration. CO and SO₂ are fairly common harmful gases. According to the standard for the allowable concentration of harmful substances in the air in the laboratory environment, SO₂ must not exceed 2 ppm and CO must not exceed 35 ppm. This research aims to establish a gas sensor array. Deep learning is a part of artificial intelligence, which has the characteristics of automatically extracting data features and solving problems. This research used artificial intelligence to calibrate gas sensor arrays. Firstly, a single gas concentration is judged. A single gas of SO₂ or CO was prepared with conditions of different temperatures and humidities. The sensor response was recorded to establish an artificial intelligence database as training and test data. Using the calibration provided by the original factory, the average SO₂ error is 2.58 ppm and the CO average error is 2.06 ppm under mixed gas. Experiments showed that the average test error of a single SO₂ sensor calibration model is 1.5 ppb, the average test error of a single CO sensor calibration model is 19.8 ppb. Calibration provided by the original factory is worse than AI model. Then a gas sensor array was set up to detect, mixed gases of SO₂ and CO with different temperatures and humidity. By recording the sensor responses, a database as training and test data was established. Using the deep neural network to calibrate and find a better model structure and data acquisition time, and selected the better performing model and acquisition length. The average test error of SO₂ of the model is 8 ppb and the average test error of CO is 166 ppb. Then used neural network modularization and kernel trick function to build a stronger model. The average test error of SO₂ of the enhanced model can be reduced to 1.7 ppb, and the average test error of CO can be reduced to 77 ppb. Using the calibration provided by the original factory, the average SO₂ error is 2.65 ppm and the CO average error is 2.69 ppm under mixed gas. Calibration provided by the original factory is worse than AI model.Finally, it is the calibration of individual differences in gas sensors. First, a universal SO₂ sensor calibration model is established, and then a gas sensor array model that can change SO₂ sensors is established. After normalizing the measurement data of four different SO₂ sensors, the model was trained and tested by a SO₂ sensor data that had never been trained before. The test error was 4.35 ppb. Combine the universal SO₂ sensor calibration model and the gas sensor array calibration model to form a gas sensor array model that can replace the SO₂ sensor, using the mixed gas data measured by the new SO₂ sensor for testing. The SO₂ average test error is 109 ppb and the CO average test error is 2450 ppb, which successfully reduces the error, and this model can replace the SO₂ sensor. Its versatility is quite helpful for the needs of a wide range of gas sensor arrays. In general, artificial intelligence can improve the accuracy of consumer-grade gas sensors, which can detect harmful gases in the environment and achieve the purpose of monitoring air quality.