以支援向量法進行不同儀器間近紅外光光譜標準化之研究

Standardization of NIR Spectra between Different Instruments by Support Vector Machines

吳志宏(Chih-Hung Wu)

指導教授 ： 陳世銘

國立臺灣大學/生物資源暨農學院/生物產業機電工程學研究所/碩士(2008年)

https://doi.org/10.6342/NTU.2008.10662

摘要

近年來，由於近紅外光（Near Infrared, NIR）技術的成熟，已建立近紅外光光譜和許多材料成份值間的關連性，因而廣泛地應用在許多的研究以及工業的領域。然而不同的近紅外光分光光度計（Spectrophotometer）所量測的光譜差異，會使得已建立的檢量線（Calibration）以及資料庫（Database）無法充分利用，此一缺點可以經由光譜標準化（Spectral standardization）加以解決。在本研究中，有別於以往線性的光譜標準化方法「片段直接標準化（Piecewise Direct Standardization, PDS）」，本研究所使用的是非線性標準化方法「支援向量標準化（Support Vector Standardization, SVS）」，乃以統計學習理論為基礎的支援向量迴歸（Support Vector Regression, SVR）所發展出來的，可以有效改進以往光譜標準化方法中無法有效處理儀器間非線性光譜差異的問題。

　　本研究採用非線性的最小平方支援向量迴歸（Least Squares Support Vector Regression, LS-SVR）進行檢量線的建立，可以獲得較佳的預測結果。為了進一步驗證SVS光譜標準化的能力，在此使用了兩組樣本光譜數據：其中第一組為文獻上常用的標準光譜數據，用來進行SVS以及以往所常被使用的光譜標準化方法PDS在光譜標準化能力的比較；第二組光譜數據為本研究實驗之粉末樣本光譜數據，將使用在判測實際情況中不同儀器間所產生的光譜差異，包括不同儀器構造以及相異的量測模組。在評斷光譜標準化的能力上，在此所使用的是光譜重建誤差（Spectral Reconstruction Error, SRE）以及標準化後的光譜預測能力。

　　利用LS-SVR所建立的檢量線皆獲得良好的預測結果；而在標準化能力的比較上，SVS皆相對於PDS獲得較佳的光譜重建能力以及標準化後光譜的預測能力。以使用相異量測模組所測量的粉末樣本之結果而言，利用LS-SVR所建立的檢量線在預測結果中，其相對之校正標準誤差（Relative Standard Error of Calibration, RSEC）以及相對之預測標準誤差（Relative Standard Error of Prediction, RSEP）分別為4.390%以及8.638%.。在光譜重建能力的比較上，校正樣本以及預測樣本光譜在使用PDS後，光譜重建誤差分別為0.0339和0.0451；而使用SVS後則各為0.0205和0.0245。相對應於未標準化前的光譜誤差分別為0.5299以及0.5227，在使用PDS與SVS後皆能大幅減少光譜的誤差，而SVS則較PDS尤佳。在標準化後光譜的預測能力比較上，未標準化前光譜的預測結果，其RSEC以及RSEP分別為134.840%以及123.670%；而在使用PDS後，其RSEC以及RSEP分別為42.355%和22.485%；使用SVS後則各為18.061%和21.441%，經標準化方法PDS以及SVS後，其光譜的預測能力皆大幅提升。

關鍵字

近紅外光 ； 光譜標準化 ； 支援向量迴歸 ； 支援向量標準化 ； 片段直接標準化 ； 最小平方支援向量迴歸

並列摘要

Recently, due to the advance of the Near Infrared technology, the relationships between NIR spectra and various components of materials have been established and generally used in industrial and academic areas. The differences of spectra among different spectrophotometers prevent the effective utilization of existed calibration models and database; however, the above-mentioned drawback can be overcome by spectral standardization. In the study, apart from the linear spectral standardization methods, such as Piecewise Direct Standardization (PDS), a non-linear method, Support Vector Standardization (SVS) which was developed based on Support Vector Regression (SVR) from statistical learning theory, was used. Support Vector Standardization can deal with the non-linear spectral differences among different spectrophotometers more effectively than conventional methods.

　　The development of calibration models in this study used the non-linear method Least Squares Support Vector Regression (LS-SVS) to obtain the better results of prediction. In order to verify the standardization capabilities of SVS, this study used two spectral data sets: the first set was the standard spectral data set available in the literature, which used to compare the past commonly used method PDS with SVS on standardization capabilities; the second one was the spectral data set of powder samples experimented in this study, which used to investigate the spectral differences made from different spectrophotometers in practice, in which the same instrument configuration equipped different module and different instrument configurations were used. In judging the standardization capabilities of PDS and SVS, the Spectral Reconstruction Error (SRE) for spectral reconstruction capability and prediction errors for prediction capability were adopted after standardization procedures.

　　As results showed, the calibration models developed by LS-SVR gained good prediction; and after standardization, SVS had better spectral reconstruction and prediction capability than those by PDS. In the case of same instrument configuration equipped different module with powder samples, the prediction result of calibration model developed by LS-SVR, the Relative Standardization Error of Calibration (RSEC) and the Relative Standardization Error of Prediction (RSEP) were 4.390% and 8.638% respectively. In the results of spectral reconstruction capability, the SRE by using PDS for calibration and prediction sets were 0.0339 and 0.0451 respectively; and when using SVS, SRE errors were 0.0205 and 0.0245 for calibration and prediction. Comparing with unstandardization case, SRE errors were 0.5299 and 0.5227. It was obvious that the SRE errors were reduced substantially after standardization; and SVS had better performance than PDS did. Regarding the results of prediction capability, before standardization, the RSEC and RSEP were 134.840% and 123.670% respectively; the RSEC and RSEP were 42.355% and 22.485% after using PDS, and they were 18.061% and 21.441% after using SVS. After standardization, RSEC and RSEP errors were reduced extensively and SVS gave better results than PDS did.

並列關鍵字

Near Infrared ； Spectral Standardization ； Support Vector Regression ； Support Vector Standardization ； Piecewise Direct Standardization ； Least Square Support Vector Regression

參考文獻

伍志翔、陳世銘、楊宜璋、陳加增、葉冠宏。2005。近紅外光光譜儀標準化模式建立之研究。出自“ 2005農機與生機論文發表會論文摘要集 ”，333-334。台北：中華農業機械學會。

伍志翔、陳世銘、楊宜璋、陳加增。2006。應用支援向量迴歸建立近紅外光光譜標準化模式。出自”2006年生物機電工程研討會論文集”，373-378。台北：台灣生物機電學會。

伍志翔。2007。近紅外光譜標準化模式之研究。碩士論文。台北：台灣大學生物產業機電工程學研究所。

歐陽孚、陳世銘、伍志翔、楊宜璋、陳加增。2006。近紅外光光譜標準化模式校正樣本篩選法則之研究。出自”2006年生物機電工程研討會論文集”，898-902。台北：台灣生物機電學會。

Belousov, A. I., S. A. Verzakov and J. von Frese. 2002. A flexible classification approach with optimal generalisation performance: support vector machines. Anal. Chim. Acta. 64(1): 15–25.