利用多標籤分類器實現電子鼻混合氣體識別方法之研究

Multi-label Classification for Mixed Odor Recognition using an Electronic Nose System

楊廷然(Ting-Ran Yang)

指導教授 ： 劉奕汶

國立清華大學/電機資訊學院/電機工程學系/碩士(2014年)

https://doi.org/10.6843/NTHU.2014.00585

摘要

目前在人工嗅覺領域中，利用電子鼻系統對混合氣體作有效的分析與識別仍是一大課題。當化學感測器陣列同時截取到多種目標氣體及當下環境的資訊時，如何採取適當的訊號處理使電子鼻系統能對抗不同的干擾，例如環境變量，背景氣體變化，感測器飄移等等，以獲取有意義的多變量響應作為特徵進行氣味識別(分類)和濃度估計(回歸)的任務，藉此增加整個系統的強健性，為本論文所關注的重點。首先，針對過去常採用電阻變化比例並取其穩態值作為特徵，本論文提出以電導變化差來進行比較，並且採用指數移動平均的概念在訊號未達穩態時擷取暫態響應，用以描述感測器對氣體的反應速率，使資料的分類鑑別度增加。接著在辨識混合氣味時，本論文採用基於構成成分作決策的多標籤辨識方法，有別於過去將單一組合皆視為一類進行分類，此方法分別對數種已知構成成分建立二元決策模型，最後將所有決策結果綜合，即為混合氣體辨識結果。與傳統方式相比，除了有效減低分類器負擔外，在低維度時也有更高的正確率。最後，為了進一步預估混合氣體各成分的濃度估計，除了使用多元線性回歸分析外，為避免資料產生共線性問題，本論文運用偏最小平方法(PLS)進行預估，此方法也可套用在分類問題上。結果顯示以電導作為特徵時，針對兩種氣體混合的情況，也能有效估計出其濃度。

關鍵字

電子鼻 ； 多標籤分類 ； 混合氣體 ； 特徵萃取 ； 多變量分析

並列摘要

In current research of artificial olfaction, effective analysis and recognition of mixed odors is still a challenging issue. Considering the specific case of a gas sensor array exposed to multiple target gases and various background gases simultaneously, and the sensor data are subject to interference from the environment such as temperature and humidity. In this research, we intend to find out some signal processing methods to get meaningful responses as multivariate features for odor identification (classification) and concentration estimation (regression). First, the thesis compares the performance with conductance difference changes and resistance fractional changes, the results show that using of conductance can achieve better accuracy rates. Secondly, we use transient features together with steady-state features. The underlying idea is that the transient phase may include additional information concerning to the constituting gases that the steady state does not provide. For the odor identification tasks, we perform multi-label classification using so-called the Individual Constituent Decision Method (ICDM) instead of conventional multi-class classification. The results show that multi-label classification reduces the computational complexity and improves the recognition accuracy. For the concentration estimation tasks, Multiple Linear Regression (MLR) is the common approach to estimate the concentration of individual constituents. In addition, in order to avoid the collinearity problem, this thesis uses the Partial Least Squares (PLS) method to estimate the concentration, and this method can also be applied to the classification problem. The results show that we can reasonably estimate the concentration of the individual constituent with conductance responses when two analyte gases existed simultaneously.

並列關鍵字

Electronic Nose ； Multi-label Classification ； Mixed Odor ； Feature Extraction ； Multivariate Analysis

參考文獻

[18] 許柏安, 「快速混合氣體辨識方法之研究」, 國立清華大學資訊工程研究所, 碩士論文, 2012.

[19] 王家銘, 「利用樣式識別實現電子鼻肺炎偵測」, 國立清華大學電機工程研究所, 碩士論文, 2013.

[1] J. W. Gardner, Handbook of Machine Olfaction. WILEY-VCH, 2003.

[2] K. Persaud and G. Dodd, “Analysis of discrimination mechanisms in the mammalian olfactory system using a model nose,” Nature, vol. 299, pp. 352–355, 1982.

[3] H.V. Shurmer, “An electronic nose: a sensitive and discriminating substitute for a mammalian olfactory system,” IEE Proc. G Circuits, Devices Syst., vol. 137, pp. 197–204, 1990.