臺灣博碩士論文加值系統

腦電圖（Electroencephalography, EEG）目前常用在輔助診斷疾病和判斷睡眠品質，但因腦電圖訊號非常容易受到干擾，這種干擾分為外在和內在的干擾，外在干擾：如磁場和金屬干擾等，內在干擾：如眨眼和頭部移動等。因此判別其來源訊號是否為腦波（Brainwave）而不是雜訊干擾波（Artifact）是一個棘手的問題。現今解決的方式有透過軟硬體修正，例如小波轉換（Wavelet Transform）、傅立葉變換（Fourier Transform）和濾波器（Filter）等。根據文獻，目前有些相關研究是利用倒傳遞類神經網路（Back Propagation Neural Network, BP Neural Network）和輻射基底類神經網路（Radial Basis Function Neural Network, RBF Neural Network）來進行腦電圖訊號識別，這些研究利用δ波、θ波、α波和β波，四種腦波來提高訊號辨識率，其辨識結果分別為75%和92%。雖然輻射基底類神經網路（RBF Neural Network）的辨識結果較高，但這個方法有一個缺點是將所有輸入值視為相同地位，沒有考慮各個輸入值的重要性高低，也就是沒有權重（Weight）的設定。於是本論文提出使用線性整流函數（Rectified Linear Unit, ReLU），這種函數其公式較簡單只取最大值，並將倒傳遞類神經網路的隱藏層（Hidden Layer）轉換函數（Transfer Function）替換成ReLU函數。而本研究使用的類神經網路為多層感知器（Multilayer Perceptron, MLP）架構，就有各層（Layers）之間的權重修正，也就有考慮到各個輸入值的重要性高低。實驗設計先經由訓練調整初始權重，使類神經網路降低誤差和改善收斂效果，以增強抗干擾的功能，再來測試是否為腦電圖訊號，統計最後的辨識率結果，得到最好的組合辨識率為93.89%，能夠減少因干擾造成辨識率的降低。

Electroencephalography（EEG）is often used to assist in diagnosing brain diseases and sleep disorders. However, EEG signals are easily interfered. The interferences can be divided into external interferences and internal interferences. For example, the former ones are caused by magnetic field or metal interferences while the latter ones are from eye blinks or head movements. Therefore, it is difficult to tell whether a signal is from a brainwave or an artifact. The current solution is to correct the signals by software or hardware, like Wavelet Transform, Fourier Transform and Filters. According to present documents, some studies have used BP and RBF neural network to identify EEG signals. δ wave, θ wave, α wave and β wave, these four types of brainwaves are adopted to discriminate the sources of the signals. They can heighten the ratio of discrimination. If used BP, the ratio can be up to 75%. If used RBF, it can be up to 92%. Although the ratio of RBF is higher, there is one disadvantage that is RBF treats all inputs equal and does not consider the priority and value of each item, meaning no set-up for weight. For this reason, in my study, the transfer function in the hidden layer of BP neural network was replaced with a ReLU function whose formula is very simple and can be used to find the maximum value. The main-used method in this study was a Multi-layer Perceptron Neural Network, in which can weight each layer, meaning to value the priority of each input item. In the experiment, through training, the weight was adjusted to lower the error and inaccuracies in Neural Network and to improve the convergent effects, all for reinforcement against interferences. Finishing the procedures above, we then tested them to make sure if they were EEG signals. The best discriminant ratio we got was 93.89%. As the result of the experiment, we could avoid lower ratio of discrimination caused by the interferences.

摘要 i
ABSTRACT ii
誌謝 iii
目錄 iv
圖目錄 vi
表目錄 vii
附錄 viii
第一章 緒論 1
1.1 前言 1
1.2 研究動機與目的 1
1.3 論文章節架構 2
第二章 相關文獻及背景介紹 3
2.1 腦波（Brainwave） 3
2.1.1 腦電圖（Electroencephalography, EEG） 4
2.1.2 腦波干擾源 5
2.2 類神經網路（Neural Network） 6
2.2.1 類神經網路簡介 6
2.2.2 監督式學習（Supervised Learning） 7
2.3 倒傳遞類神經網路（Back Propagation Neural Network, BP Neural Network） 9
2.3.1 腦波能量資料前處理 10
2.3.2 雙彎曲函數（Sigmoid Function） 11
2.3.3 均方差（Mean Square Error, MSE） 11
2.3.4 梯度下降法（Gradient Descent） 12
2.4線性整流函數（Rectified Linear Unit, ReLU）及各層（Layers）的修正 13
2.4.1 線性整流函數 13
2.4.2 隱藏層（Hidden Layer）至輸出層（Output Layer）之修正 15
2.4.3 輸入層（Input Layer）至隱藏層（Hidden Layer）之修正 16
2.5 類神經網路訓練和測試方式 18
2.5.1 訓練方式 18
2.5.2 測試方式 19
2.6 相關論文探討 20
第三章 使用類神經網路增強腦電圖訊號抗干擾功能 23
3.1 問題描述 23
3.2 研究設備 23
3.3 研究設計方式 25
3.4 本實驗類神經網路詳細訓練步驟和測試步驟 31
3.4.1 訓練步驟 31
3.4.2 測試步驟 32
3.5 研究架構流程 33
第四章 實驗結果與數據分析 34
4.1 實驗受試者之資料 34
4.2 腦波能量擷取 35
4.3 類神經網路訓練過程和結果 36
4.3.1 最初訓練結果 37
4.3.2 中間訓練結果 38
4.3.3 最終訓練結果 39
4.4 類神經網路測試結果 42
第五章 結論 45
參考文獻 46
附錄 49

[1]A.Zhang, “EEG Feature Extraction and Analysis under Drowsy State Based on Energy and Sample Entropy,” 2012 5th Int. Conf. Biomed. Eng. Informatics (BMEI 2012), no. 30670529, pp. 501–505, May. 2012.
[2]Y.Liu, L.Yan, B.Zeng, and W.Wang, “Automatic sleep stage scoring using hilbert-huang transform with BP neural network,” 2010 4th Int. Conf. Bioinforma. Biomed. Eng. iCBBE 2010, no. 30970876, pp. 90–93, June. 2010.
[3]Y.Tomita, F. B.Vialatte, G.Dreyfus, Y.Mitsukura, H.Bakardjian, and A.Cichocki, “Bimodal BCI using simultaneously NIRS and EEG,” IEEE Trans. Biomed. Eng., vol. 61, no. 4, pp. 1274–1284, Apr. 2014.
[4]Z.Ni and N.Malla, “Towards Enabling Deep Learning Techniques for Adaptive Dynamic Programming,” 2017 Int. Jt. Conf. Neural Networks, pp. 2828–2835, May. 2017.
[5]X.-B.Qin, Y.-Z.Zhang, M.-T.Huang, and M.Wang, “EEG Signal Recognition Based on Wavelet Transform and Neural Network,” 2016 Int. Symp. Comput. Consum. Control, pp. 523–526, July. 2016.
[6]X.Glorot, A.Bordes, andY.Bengio, “Deep sparse rectifier neural networks,” AISTATS ’11 Proc. 14th Int. Conf. Artif. Intell. Stat., vol. 15, pp. 315–323, Apr. 2011.
[7]A.Krizhevsky, I.Sutskever, and H.Geoffrey E., “ImageNet Classification with Deep Convolutional Neural Networks,” Adv. Neural Inf. Process. Syst. 25, pp. 1–9, Dec. 2012.
[8]G. Deuschl and A. Eisen, “Guidelines of the International Federation of Clinical Physiology (EEG Suppl. 52),” 1999 Int. Fed. Clin. Neurophysiol., Mar. 1999.
[9]G.Klem, H.Luders, H.Jasper, and C.Elger, “The ten-twenty electrode system of the International Federation,” Electroencephalogr. Clin. Neurophysiol., vol. 10, no. 2, pp. 371–375, Mar. 1999.
[10]J.Malmivuo andR.Plonsey, Bioelectromagnetism. OXFORD UNIVERSITY PRESS, Jan. 1995.
[11]Robert Hecht-Nielsen, Neurocomputing. Addison-Wesley Publishing Company, Feb. 1990.
[12]Z.Christoph, T.Johannes, and Z.Carl, Brain-state dependent brain stimulation: Real-time EEG alpha band analysis using sliding window FFT phase progression extrapolation to trigger an alpha phase locked TMS pulse with 1 millisecond accuracy. Dec. 2015.
[13]R.Hecht-Nielsen, “Theory of the backpropagation neural network,” in Neural networks for perception (Vol. 2), Harcourt Brace & Co. Orlando, FL, USA, Jan.1992, pp. 65–93.
[14]Jun HanClaudio Moraga, “The influence of the sigmoid function parameters on the speed of backpropagation learning,” From Nat. to Artif. Neural Comput. Int. Work. Artif. Neural Networks, vol. 930, pp. 195–201, June. 1995.
[15]S.Masood, M. N.Doja, and P.Chandra, “Analysis of weight initialization techniques for Gradient Descent algorithm,” 12th IEEE Int. Conf. Electron. Energy, Environ. Commun. Comput. Control (E3-C3), INDICON 2015, no. l, Dec. 2015.
[16]R. and al.Douglas, “Recurrent excitation in neocortical circuits,” Science (80-. )., p. 269(5226), 981-5, Aug. 1995.
[17]R.L.T. Hahnloser, “On the piecewise analysis of networks of linear threshold neurons,” Neural Networks, vol. 11, no. 4, pp. 691–697, June. 1998.
[18]P.Lennie, “The Cost of Cortical Computation,” Curr. Biol., p. Vol. 13, Iss. 6, 493-497, Mar. 2003.
[19]X.Glorot and Y.Bengio, “Understanding the difficulty of training deep feedforward neural networks,” Proc. 13th Int. Conf. Artif. Intell. Stat., vol. 9, pp. 249–256, May. 2010.
[20]V.Nathan and R.Jafari, “Reducing the noise level of EEG signal acquisition through reconfiguration of dry contact electrodes,” IEEE 2014 Biomed. Circuits Syst. Conf. BioCAS 2014 - Proc., pp. 572–575, Oct. 2014.
[21]劉勇均, “癲癇腦波訊號之棘波偵測與辨識,” 博士論文, 國立成功大學資訊工程學系, Feb. 2014.
[22]彭仁平, 董東璟, 陳憲宗, 蔡政翰蔡仁智, “應用類神經網路建置瘋狗浪預警系統,” 第38屆海洋工程研討會暨科技部計畫成果發表會, pp. 65-70, Dec. 2016.
[23]陳昌捷, “以倒傳遞類神經網路預測股市指數,” 碩士論文, 國立宜蘭大學多媒體網路通訊數位學習碩士在職專班, Aug. 2015.
[24] 湯雅雯, “腦波量測系統之研製與腦波信號之非線性分析,” 碩士論文, 國立成功大學電機工程學系, July. 2004.
[25] 葉怡成, 類神經網路模式應用與實作. 儒林圖書股份有限公司, pp. 4-14, Apr. 2009.
[26] 張斐章;張麗秋, 類神經網路導論：原理與應用. 滄海書局, pp. 85, Jan. 2010.