- 學位論文
近紅外光光譜標準化模式建立之研究
Study on the Development of the Standardization of Near-Infrared Spectra
摘要
本研究以PDS (Piecewise Direct Standardization)、DPDS (Differential Piecewise Direct Standardization)及SVS(Support Vector Standardization)三種方法針對因為儀器的不同造成光譜在強度上的差異作標準化，並且討論標準化後光譜應用於已建立之資料庫的準確度。標準化方法PDS及DPDS是以PLSR為基礎之線性標準化的方法，而SVS是以SVR為基礎非線性標準化方法，另外為提升標準化光譜預測標準參考值(如水果糖度)之能力，以KS所選別之標準化校正樣本，少量但是差異性較大之樣本對已建立之資料庫作Model Adaptation，使標準化光譜可用性更高。蔗糖(Sucrose)與鹽(NaCl)的混合物為本研究之實驗樣本，以其作光譜量測並以蔗糖含量作為標準參考值，分佈範圍為0%~100%。採用PDS、DPDS及SVS三種方法對不同時間量測、相同儀器不同配件、不同儀器之光譜作標準化，分別討論光譜誤差與標準參考值之預測誤差。分析結果發現以SVS作標準化後之結果最佳。不同時間量測之光譜標準化其光譜誤差SEPw=0.044，既有資料庫經過Model Adaptation後，標準參考值預測之誤差SEPc=6.464。相同儀器不同設備的實驗，NIRS 6500之樣本夾具mini cup與fiber配件量測後標準化之光譜誤差SEPw =0.050，以既有之資料庫作預測，標準參考值誤差SEPc=12.911，資料庫經過Model Adaptation後SEPc=6.471。不同儀器的實驗，Online 6500與NIRS 6500標準化後之光譜誤差SEPw = 0.040，以既有之資料庫作預測，標準參考值誤差SEPc=8.482，Model Adaptation後，標準參考值預測之誤差SEPc=7.318；N-400與NIRS 6500兩台儀器標準化後之光譜誤差SEPw=0.042，資料庫經過Model Adaptation後標準參考值預測之誤差為SEPc = 5.623。Online 6500與N-400兩台儀器波長範圍及間距不等，若不作光譜前處理(波段選擇及波長間距調整)，標準化後之光譜誤差SEPw = 0.034，既有資料庫經過Model Adaptation後誤差SEPc=6.03，而經過波段選擇1100~2000nm後光譜誤差SEPw=0.02，已既有之資料庫作預測誤差SEPc=7.800。另外本研究亦討論當儀器標準化主副關係互換，結果發現當誤差較大精確度較差之儀器轉換到精確度較高的儀器時，光譜誤差較大，標準參考值之預測精準度也隨之降低，Online 6500及NIRS 6500互換後標準化光譜誤差SEPw=0.072，已既有之資料庫作預測SEPc=15.338，若選擇波段以1100~2000nm作標準化，可提升標準化能力，光譜誤差SEPw=0.016，已既有之資料庫作預測SEPc=12.470。由結果來看，本研究所研發出之光譜標準化方法可成功地將不同時間測量、兩台不同設備或不同儀器量測光譜所造成的差異降低，並且透過Model Adaptation的方法提升標準化光譜應用於既有資料庫的預測能力。
並列摘要
This research used PDS (Piecewise Direct Standardization), DPDS (Differential Piecewise Direct Standardization) and SVS (Support Vector Standardization) to standardize spectra, which were obtained by using different instruments. Among them, PDS and DPDS are linear standardization methods based on PLSR (Partial Least Square Regression), while SVS is a nonlinear standardization method based on SVR (Support Vector Regression). Besides, to improve the use of standardized spectra on established database and models, Model Adaptation was introduced with samples were selected by KS (Kennard-Stone algorithm). The mixture of sucrose and NaCl was used as samples in this study, in which the content of sucrose was served as the reference data. Three models of PDS, DPDS and SVS were adopted to investigate the spectra standardization for the cases of different time, attachments and instruments. The results indicated that SVS was the best model. Regarding the effect of measurements at different time, spectra error SEPw was 0.044 after standardization, and standard refrence value error SEPc was 6.464 after Model Adaptation. Regarding the measurements by using different attachments on the same instrument, SEPw and SEPc were 0.050 and 12.911 respectively after standardization, and SEPc was 6.471 after Model Adaptation. As of the measurements by using different instruments (Online 6500 and NIRS 6500), SEPw and SEPc were 0.040 and 8.842 respectively after standardization, and SEPc was 7.318 after Model Adaptation. As of the measurements by using different instruments (N-400 and NIRS 6500), spectra error SEPw was 0.042 after standardization, and standard refrence value error SEPc was 5.623 after Model Adaptation. As of the measurements by using different instruments (N-400 and Online 6500), spectra error SEPw was 0.034 after standardization, and standard refrence value error SEPc was 6.03 after Model Adaptation; however, if same wavelength range (1100~2000nm) was selected, standard error of spectra SEPw was 0.02 and SEPc was 7.800 without Model Adaptation. In this study, the investigation was also conducted by exchanging the relation of Master and Slave in the standardization. The results revealed that when a less precise instrument was standardized to a more precise instrument, both SEPw and SEPc increased. After exchanging of Online 6500 and NIRS 6500, SEPw and SEPc were 0.072 and 15.338 respectively after standardization; if same wavelength range (1100~2000nm) was selected, standard error of spectra SEPw was 0.016 and SEPc was 12.470. Spectra standardization methods with Model Adaptation were successfully developed in this study, it allowed to transfer spectra among different attachments and instruments, and to reduce the spectra difference due to different measurement time.