- 學位論文
P300腦機介面拼字器之線上校正系統開發
Development of an online calibration system for P300 brain-computer interface Speller.
摘要
在1988年，Farwell與Donchin首次提出一套腦機介面(Brain-Computer Interface,BCI)系統－P300 Speller，並成功的利用視覺刺激面板誘發出P300事件相關電位。然而，此系統發展之目的在於改善病患生活品質，使病患具有與外界溝通的能力，而各種領域之研究學者也紛紛投入此研究，以發展出精度高且速度快之P300拼字器為目的。故本論文提出具有校正機制之P300拼字系統，此系統達到快速訓練分類模型以及快速蒐集資料的特點。受測者在蒐集腦波資料後，藉由疊加平均原理將P300事件相關電位之特徵突顯出來，再進行本論文所設計之線上校正系統，即可改善P300拼字器蒐集大量訓練資料以及訓練支持向量機器之分類模型的時間。本系統分為兩種方法，第一種方法每位受測者皆使用自己的腦波資料，利用支持向量機器建立分類模型，並以自己的模擬線上資料進行測試。第二種方法為具有校正機制之P300拼字器，此機制將既有之初始資料庫，加入特定受測者所提供之校正資料庫，將兩種資料合成為延伸資料庫，進行漸進學習式支持向量機器快速的建立分類模型。第一種方法在模擬線上資料之最高分類率雖然達到99.39%，但在支持向量機器的訓練複雜度非常高，導致訓練時間過長。然而，具有校正機制之結果顯示，利用漸進學習式支持向量機，其建立分類模型之時間僅需約一分鐘，且分類結果最高達到98.17%。在本論文中加入三種特徵抽取方法進行比較，在分別為主成份分析( Principal Component Analysis, PCA)、核心主成份分析( Kernel Principal Component Analysis, KPCA)、線性識別分析(Linear Discriminant Analysis, LDA)。其結果顯示，將資料經由PCA轉換後其具有校正機制之系統在使用漸進學習式支持向量機器的情況下，最高分類率達到99.31%，其中有六位受測者之分類率提升了5%~6%，且最快運算時間僅需11.99秒即可完成分類模型。本篇論文設計之線上校正系統使P300拼字器具有快速且精度高之特性，這將大幅提升了未來病患在使用P300線上拼字系統的效益。
並列摘要
In 1988, Farwell and Donchin first proposed a Brain-Computer Interface system (BCI system) – P300 Speller, and successfully used the vision-stimulation panel to evoke the P300 event related potential. The goal of developing this system was to help improve the living quality of patients by enhancing their capability of communicating with others. A lot of researchers also dedicated to this research, and took developing the P300 spelling with high classification rate and quick speed as the goal. In this paper, we proposed a tuning scheme of P300 Speller system. This system possesses the features of fast data-collection and fast training model. The subjects collect the EEG data, and then highlight the characteristics of the P300 event related potential using the principle of superposition mean as well as the online tuning scheme designed in this paper; in this way, the time of P300 speller collecting a large amount of training data and training of SVM classification models can be reduced. The system includes two modes. The first mode, the subjects use their own EEG data to build the SVM classification models, and then use the data for the online simulation tests. The second mode is the tuning scheme of the P300 Speller. This scheme combines the original data base with the corrected data base provided by certain subjects into the extended data base. This extended data base is used to reduce the time of model-creation by using the incremental learning SVM method. Although the highest accuracy of the first mode reaches 99.39% in the simulation online data, its training complexity of SVM is relatively high, and thus the training time would be too long. On the other hand, the results produced by the system with tuning scheme show that the time of model-creation only takes one minute by utilizing the incremental learning SVM method, and the best result can reach 98.17%. Three feature-extraction methods are put into comparison in this paper: Principal Component Analysis (PCA), Kernel Principal Component Analysis (KPCA), and Linear Discriminant Analysis (LDA). The results show that, after transforming the data via PCA, under the condition that the tuning scheme of system utilizing the incremental learning SVM method, the highest classification rate reaches 99.31%. The classification rates of six out of the subjects increase by 5% to 6%, and the shortest computation time of creating the classification model only takes 11.99 seconds. The online tuning scheme of P300 system designed in this paper owns the features of high speed and high accuracy and these will greatly improve the benefits of the patients for the using of the online P300 spelling system in the future.