希爾伯特黃轉換應用於單筆腦電波訊號

Abstract

大腦人機介面為近幾年來很活躍的研究領域，然而腦電波包含著許多雜訊，並且具有非線性、非穩態等特性，所以腦電波的「特徵擷取」和「分類」為世界各國相關研究團隊共同努力的方向。本研究針對「特徵擷取」提供一套單筆腦電波分析法，可應用於大腦人機介面。此方法同時結合了獨立成份分析法和希爾伯特黃轉換來分析腦電波訊號。整個分析方法可以分為兩個階段：第一階段是利用獨立成份分析法的盲源分離特性把腦電波紀錄分解成具有時間、空間特性的成分；第二階段是利用一種時頻分析法，希爾伯特黃轉換將第一個階段所得到的成分做時頻分析。其中希爾伯特黃轉換的前半段架構被稱為經驗模態分解法，在本論文中用於訊號重建以及雜訊去除。當腦電波經過前面兩個階段的處理後，將可以同時保留腦電波的時間、頻率、空間特徵，建立時頻空間樣本。另外，本研究同時比較三種時頻分析法在處理腦電波訊號上的表現，分別為希爾伯特黃轉換、短時距傅立葉轉換、莫雷轉換。實驗結果為希爾伯特黃轉換在頻譜圖上不論高頻和低頻皆有較佳的解析度。 在研究流程的最後一個階段，選擇支持向量機分類四種想像動作(想像左手動、想像右手動、想像腳動、想像舌頭動)的腦電波資料，對於想像右手動可到達83.33%的辨識率，而整體四種想像動作的平均辨識率為54.17%。另外，本研究結合經驗模態分解與快速獨立成份分析法定位想像動作之訊號源，平均準確率可達到72.12%。Brain computer interface (BCI) is a highly active field in the recent years. However, electroencephalogram (EEG) is a non-linear and non-stationary signal which contains various noises. Hence, the research groups in the world put effort into “feature extraction” and “classification” of EEG signal. In this study, an analysis strategy of single-trial EEG is developed for BCI and focus on “feature extraction”. The proposed method is to combine Independent Component Analysis (ICA) and Hilbert-Huang Transform (HHT) to analyze EEG. The EEG signal processing is performed in two stages: (1) ICA, a kind of Blind Source Separation (BSS), is used for dividing temporal-spatial components of EEG recordings; (2) HHT, a kind of temporal-frequency analysis (TFA), is used for processing these components. The empirical mode decomposition (EMD) in part of HHT is used for signal reconstruction and noise rejection. When the EEG is processed by above mentioned methods, the time-frequency-spatial features will be preserved. In additional, three kinds of TFAs (HHT, short time Fourier transform (STFT), and Morlet Transform) are experimented to compare the performance. The results show that the HHT has the better resolution in high frequency and low frequency. The last stage of the research, Support vector machine (SVM) is selected to classify EEG data of four motor imageries (left hand, right hand, foot, and tongue). The experiment result presents that the classification accuracy of imaging “right hand” moving is 83.33%, and total averaged accuracy is 54.17%. In additional, the method combined with both EMD and FastICA has 72.12% averaged accuracy of signal source localization of motor imagery.