目前電子鼻系統對於單一氣體辨識的演算法,已具有一定的辨識水準,但若延伸至一般生活情境的混合氣體辨識應用,則仍有許多進步的空間。現有之混合氣體辨識的演算法研究,目的在於開發出具低複雜度的演算法、或者是將混合氣體中的未知成份予以排除,不影響分類結果。此外,無論是單一或混合氣體辨識的研究,皆需要等待感測器的氣體響應訊號達到恆定態,以取其恆定值作特徵,並且往往是於理想的實驗環境下感測氣體;而在生活中,這樣的等待時間會帶來不便,現實中也沒有所謂的理想環境,因此不符合實際應用。本論文提出一個新的觀點-混合氣體訊號在未達恆定態前亦具有分類鑑別度的資訊。本論文在縮短實驗時間以及於非理想實驗環境感測氣體的兩項前提下,提出一演算流程: 依此對訊號作前處理,以及除了定時間的響應值外,再加入氣體響應訊號的反應速率與其出現的時間位置當作特徵;之後透過特徵選取的技巧取得關鍵特徵後,再使用分類器做辨識。結果顯示,與未經處理、實驗時間最久的90%辨識率相比,透過我們的演算流程,提早1/4時間處辨識混合氣體,辨識率從65%提升至82%;提早1/2時間處,辨識率從83%提升至87%,顯著提升提早辨識的效果。
For electronic nose systems, existing algorithms for single odor analysis have provided commercial-grade recognition results. However, for practical applications to deal with mixed odors, effective algorithms are yet to be developed. One of the problems with mixed-odor analysis using metal-oxide semiconductor sensors is the long saturation time (to reach steady-state) of sensor responses which are important features for odor analysis. In this thesis, we propose an efficient method to reduce the time to recognize mixed odors before the sensor responses reach saturation states. Our method consists of signal pre-processing, sensor response feature extraction, feature selection and normalization, and K-Nearest Neighbor (KNN) classification. Experiments show that the proposed method improves mixed-odor analysis time (using metal-oxide semiconductor sensors) significantly without sacrificing the accuracy of the recognition.