光譜理論發展至今已有三百多年的歷史，而藉此發展的光譜儀則廣泛地運用在天文、衛星遙測、醫藥、工業等光譜量測用途上。然而，光譜訊號容易因量測儀器之硬體或量測環境所影響，產生不可預期的雜訊。
本論文主要在探討光譜儀量測訊號之處理，採用單一個別波長訊號資料做即時地動態移動平均，有別於一般光譜儀通常採用固定平均及曲線平滑的作法。本研究之光譜訊號處理分為三大核心模組，包括類神經網路模組、移動平均階度決策模組及動態移動平均模組。其中類神經網路透過離線學習建模，進行時間序列的訊號分析，提供推估基準值給移動平均階度決策模組，根據決策法則與單一波長實際值之差值比較推算出合適之移動平均階度。之後動態移動平均模組再根據此階度即時調整階度，與計算全域之個別波長動態移動平均值，達成以較低的移動平均次數而提昇訊號雜訊比的目的。
而決策模組之關鍵在於當實際與推估之差值百分比達到設定門檻時，模組主動將該筆資料剔除，並以前次之平均階度取代，可大幅降低移動平均之次數。實驗以一組200筆光譜儀量測資料做代表，證實以較低的動態移動平均次數做訊號處理後，其SNR值比採用固定移動平均次數要優異，如與固定移動平均8次做比較，動態移動平均6.095次，平均次數少了23.81%，且訊號的SNR卻高出7.47%。

Spectrum theory has been developed over the past 300 years. Modern spectrometers are now widely applied in astronomy, satellite telemetry, medical science, industry, etc. However, the spectral data is easily affected by hardware and environmental conditions.
This study presents a spectrometer signal processing method which is different from traditional moving average (MA) and curve smoothing methods. The new method individually processes the data for a single wavelength with dynamic moving average. Three major modules are proposed: a neural network (NN) module, a moving average decision module, and a dynamic moving average module. During the signal processes, the modeled neural network performs a time series analysis on the actual values for a single wavelength and produces a predicted value. The moving average decision module compares the predicted value and actual value for the specific wavelength and produces a proper number of moving average according to the given rules. The dynamic moving average module then uses this number of moving average to calculate the value of moving average individually for each single wavelength in the spectrum. In such a way, it is able to improve the signal-to-noise ratio (SNR) with lower average numbers of moving average.
The key for reducing the average number in the moving average decision module is the given threshold that is used to determine whether a noise data is removed or not. The removed data is replaced with previous data and the number of moving average is kept as the previous value. The experiment uses a set of data with 200 samples and shows that the SNR by dynamic moving average method results in fewer calculation steps and higher SNR value. For such a data set, the dynamic moving average method gave an average of 6.095 calculation steps, 23.81% less than the fixed moving average method of 8 steps, and the SNR value for the proposed method shows an improvement of about 7.47%.

參考文獻
[1]Encyclopædia Britannica Online , prism: Newton’s experiment, 2011, http://www.britannica.com/EBchecked/media/1066/Newtons-prism-experiment
[2]M. H. Freeman, C. C. Hull, Optics, 2003.
[3]國立臺灣師範大學物理學系網站，「可見光」，2011。
[4]江月松、宋馭民、劉正瑜、全中興、陳積德，「光電技術」，vol. 初版，台北， 新文京開發出版股份有限公司，2005。
[5]劉興鑑、孫逸民、陳玉舜、趙敏勳、謝明學，「儀器分析」，初版，台北，全威圖書有限公司，2007。
[6]J. F. James, Spectrograph design fundamentals, Cambridge ; New York, Cambridge University Press, 2007.
[7]Mortimer Abramowitz , Michael W. Davidson, Numerical Aperture and Resolution, 2010,
http://www.olympusmicro.com/primer/anatomy/numaperture.html
[8]W. J. Smith, Modern Optical Engineering, 3rd ed. New York, McGraw-Hill, 2000.
[9]Shimadzu, Shimadzu UV-VIS-NIR Spectrophotometer, 2011, http://www.shimadzu.com/products/lab/spectro/oh80jt0000001k9a.html
[10]oto photonics inc， 「超微光譜儀技術 - 光學機構」，2010， http://www.otophotonics.com/big5/Technology_01.asp
[11]R. L. McCreery, Raman Spectroscopy for Chemical Analysis, New York, John Wiley & Sons, 2000.
[12]M. F. Tompsett, G. F. Amelio, W. J. Bertram, Jr., R. R. Buckley, W. J. McNamara, J. C. Mikkelsen, Jr., D. A. Sealer, "Charge-coupled imaging devices: Experimental results", Electron Devices, IEEE Transactions on, vol. 18, pp. 992-996, 1971.
[13]繆家鼎、徐文娟、牟同升，「光電技術」, 初版，台北市，五南圖書出版股份有限公司，2003。
[14]Saeed V. Vaseghi, Advanced Digital Signal Processing and Noise Reduction, 4 ed, Wiley, 2008.
[15]D. J. Schroeder, Astronomical optics, Academic Press, 2000.
[16]Robert de Levie, How to use excel in analytical chemistry and in general scientific data analysis, Cambridge, University of Cambridge, 2004.
[17]Douglas C. Montgomery, Cheryl L. Jennings, Murat Kulahci, Introduction to Time Series Analysis and Forecasting, Hoboken. New Jersey, John Wiley & Sons. Inc., 2008.
[18]H. Liu, H.-Q. Tian, C. Chen, Y.-f. Li, "A hybrid statistical method to predict wind speed and wind power", Renewable Energy, vol. 35, pp. 1857-1861, 2010.
[19]A. Sfetsos, C. Siriopoulos, "Time Series Forecasting of Averaged Data With Efficient Use of Information", Systems, Man and Cybernetics, Part A: Systems and Humans, IEEE Transactions on, vol. 35, pp. 738-745, 2005.
[20]Matin T. Hagan, Howard B. Demuth, Mark Beale, Neural Network Design, Boston MA, THOMSON, 1996.
[21]張斐章、張麗秋，「類神經網路」，台北，台灣東華書局股份有限公司，2005。
[22]Jeff Heaton, Introduction to Neural Networks for C#, 2nd ed, St. Louis, Heaton Research, Inc, 2008.
[23]Matin T. Hagan, Howard B. Demuth, Mark Beale, Neural Network Design, 初版，台北市，新加坡商湯姆生亞洲私人有限公司台灣分公司，2004。
[24]羅華強，「類神經網路：MATLAB的應用」， 新北市，高立圖書有限公司，2005。