| Abstract: | 本論文是以鈉超離子導體(NASICON)作為電解質完成電流式氣體感測器製作，進行對揮發性有機氣體之感測，並組裝為陣列式胺類氣體感測器，輔以倒傳遞類神經網路對胺類混合氣體進行預測分析。 首先製作NASICON/白金(黃金)/氧化鋁基板為感測電極，進行各種揮發性氣體之感測特性探討，感測之氣體有甲醇、乙醇、甲醛、丙醛、丙酮、丁酮、甲胺、二甲胺與三甲胺。發現應答時間由快至慢依序為胺類<酮類<醛類?醇類，回復時間則以胺類為最長，醛類與醇類次之，酮類最短。而在感測靈敏度上以甲胺在白金電極所得到的0.494nA/ppm為最大，丙酮得到0.262nA/ppm為最小。 利用微組裝技術製作NASICON/白金(黃金)/氧化鋁基板之陣列式感測器，偵測混合之胺類氣體。感測器上之二組白金電極之電位分別設定在0.85V與1.45V，黃金電極之電位則設定在1.10V，以分別對甲胺、三甲胺與二甲胺之混合氣體作感測。將所得之感測訊號以倒傳遞類神經網路進行學習訓練與模擬。在50組皆為三成分混合氣體的學習數據下，將學習架構設定為3、15、30、3，學習速率為0.3，學習循環數為500,000次，所得之學習誤差最小為3.62%。應用此結果於未知氣體的分析預測，其氣體之真實濃度與倒傳遞網路預測值間之平均誤差均小於3%。 若以混合有單成分、二成分與三成分之胺類之氣體訊號進行學習，則網路架構為3、40、80、3，學習速率為0.3，學習循環數為5,000,000次時，其學習誤差最小為2.88%。進行未知氣體之預測時，甲胺氣體與二甲胺氣體之誤差不超過3.1%與2.6%，三甲胺氣體之最大誤差為8.1%，最小誤差為1.8%。
In this thesis, the amperometric gas sensors using NASICON as electrolyte to detect the volatile organic compounds (VOCs) were studied. The concentrations of amines in the mixture gas were monitored by the sensor array in based on BPN (back-propagation network). The sensing properties for various VOCs, including methanol, ethanol, formaldehyde, propionaldehyde, acetone, methyl ethyl ketone (MEK), methylamine, dimethylamine and trimethylamine, were analyzed by NASICON/Pt(Au)/Al2O3 prepared in this work. The response time was in the order of amines < ketones < aldehydes ? alcohols, and the order of recovery time was found to be ketones < aldehydes < alcohols < amines. The maximum and minimum sensitivities were obtained as 0.494 nA/ppm and 0.262 nA/ppm for detecting methylamine and acetone on the Pt, respectively. The sensor array prepared with the microfabrication technniques was used to monitoring the mixture of amines gas. The potentials of two Pt electrodes on the sensor array were set at 0.85 V and 1.45 V to detect methylamine and trimethylamine, respectively, and the potential of Au electrode on the sensor array was fixed at 1.10 V to monitor dimethylamine. The signals from the sensor array was learned and analyzed with BPN. The optimal network structure, learning step and learning cycle were found to be (3, 15, 30, 3), 0.3 and 500,000, respectively, based on 50 learning data (3 mixture gases), and the learning error was 3.62%. Based on this optimal BPN structure, the average error for analyzing the mixture gases containing 3 amines was less than 3%. When the 50 learning data contained the data sets of the single, double and triple amines, the optimal structure, learning step and learning cycle were found to be (3, 40, 80, 3), 0.4 and 5,000,000, respectively, and the learning error was 2.88%. When this BPN structure was used to analyze the amines mixture gas the analyzing errors of methylamine and dimethylamine were found to be 3.1% and 2.6%, respectively, and the analyzing error of trimethylamine was obtained in the range of 1.8% ~ 8.1%. |
