以人工智慧結合氣體感測器陣列偵測混合氣體之種類與濃度

Detection of Gas Composition by Using Gas Sensor Array with Artificial Intelligence

張至綱(Chih-Kang Chang)

指導教授 ： 吳政忠

國立臺灣大學/工學院/應用力學研究所/碩士(2019年)

https://doi.org/10.6342/NTU201900577

摘要

空氣汙染為近年來逐漸受到重視的議題之一，其中一氧化碳(CO)氣體與二氧化硫(SO2)氣體為兩種空氣汙染指標物。一氧化碳為常見之有害氣體之一，透過燃燒不完全而產生，具有無色、無味之特性，長期超標吸入體內將會損害人體心肺功能，或是產生頭痛、暈眩甚至死亡等現象。二氧化硫也是空氣汙染來源之一，主要為工業石化燃料燃燒所生成，具有無色、不可燃、酸性與腐蝕性等特性，為造成酸雨之主要物質。而空氣中之汙染物成分與濃度必須依靠氣體感測器來進行量測與監控，然而現今氣體感測器存在氣體選擇性不佳之問題，因此本論文將針對氣體選擇性問題提出改善方法，藉由人工智慧結合氣體感測器陣列來偵測一氧化碳與二氧化硫之混合氣體之種類與濃度。
本研究將建立並使用深度類神經網路與卷積類神經網路兩種模型進行混合氣體之種類判別與濃度預測，並透過四種不同輸入數據型態作為類神經網路輸入資料來源，分別為時間域訊號、讀值偏移量、雷達圖圖像以及時間域影像。由實驗結果發現，時間域訊號、讀值偏移量、雷達圖圖像以及時間域影像之測試誤差依序為7.729、6.482、8.564、5.483 ppm，因此時間域影像作為輸入數據型態並結合卷積類神經網路模型對於混合氣體之種類判斷與濃度分析會有最佳預測結果。在資料分布方面，以氣體感測器多寡、濃度點分布以及資料數量做探討分析，當氣體感測器越多，濃度點分布間距越緊密，資料筆數越多，對於類神經網路模型之整體學習效益越佳。
最後將多元迴歸分析與人工智慧方法進行比較，其測試誤差分別為26.627與5.483 ppm，證明人工智慧適用於氣體種類判別與濃度預測分析，成功改善並解決氣體感測器之選擇性問題，實現偵測空氣中環境之變化與污染並達到監控空氣品質之目的。

關鍵字

人工智慧 ； 深度學習 ； 類神經網路 ； 氣體感測器陣列 ； 一氧化碳感測 ； 二氧化硫感測

並列摘要

Air pollution is one of the topics that has received increasing attention in recent years. Carbon monoxide (CO) gas and sulfur dioxide (SO2) gas are two air index pollutants. Carbon monoxide is one of the common harmful gases. It is produced by incomplete combustion. It has colorless and odorless characteristics. If it is inhaled in the body for a long time, it will damage the cardiovascular function of the human body, or cause headache, dizziness or even death. Sulfur dioxide is also one of the sources of air pollution. It is mainly produced by industrial fossil fuel combustion. It has the characteristics of colorless, non-flammable, acidity and corrosivity, and is the main substance causing acid rain. The composition and concentration of pollutants in the air must be measured and monitored by gas sensors. However, there are gas selectivity problems with gas sensors nowadays. Therefore, this thesis will propose an improvement method for gas selectivity problems. The type and concentration of the mixed gas of carbon monoxide and sulfur dioxide are detected by using gas sensor array with artificial intelligence.
This study will establish and use deep neural network and convolutional neural network for classifying the type of gases and predicting the concentration of mixed gas, and through four different input data types as the input source of neural network. They are time domain signal, reading value shift, radar chart, and time domain image, respectively. The experimental results show that the test errors of time domain signal, reading value shift, radar chart and time domain image are 7.729, 6.482, 8.564, 5.843 ppm, respectively, so the time domain image is used as the input data type and combined with the convolutional neural network model, there will be the best prediction results for the type decision and concentration analysis of the mixed gas. In terms of data distribution, the number of gas sensors, the concentration point distribution and the number of data are analyzed. When the gas sensors are more, the closer the concentration point distribution is, the more data is available, the overall learning benefit is better for the neural network model.
Finally, the multiple regression analysis is compared with the artificial intelligence method. The test errors are 26.627 and 5.843 ppm, respectively, which proves that artificial intelligence is suitable for gas species discrimination and concentration prediction analysis, and successfully improves and solves the selectivity problem of gas sensor, that can realize the detection of changes and pollution in the air and achieve the purpose of monitoring the air quality.

並列關鍵字

Artificial intelligence ； Deep learning ； Neural network ； Gas sensor array ； Carbon monoxide sensing ； Sulfur dioxide sensing

參考文獻

[1] 行政院環境保護署-空氣品質監測網. Available: https://taqm.epa.gov.tw/taqm/tw/default.aspx

[2] M. Belaqziz, M. b. Amjoud, A. Gaddari, B. Rhouta, and D. Mezzane, "Enhanced room temperature ozone response of SnO2 thin film sensor," Superlattices and Microstructures, vol. 71, pp. 185-189, 2014.

[3] Z. Zhu, J.-L. Chang, and R.-J. Wu, "Fast ozone detection by using a core–shell Au@ TiO2 sensor at room temperature," Sensors and actuators B: chemical, vol. 214, pp. 56-62, 2015.

[4] W. Hoeflinger and T. Laminger, "PM2.5 or respirable dust measurement and their use for assessment of dust separators," Journal of the Taiwan Institute of Chemical Engineers, 2017.

[5] L. Djoumi, V. Blondeau-Patissier, M. Vanotti, J.-C. Appert-Collin, D. Thomas, and L. Fertier, "Surface Acoustic Wave Sensors for PM2.5 and PM10 Concentration," Procedia Engineering, vol. 168, pp. 696-699, 2016.