基於無線腦機介面之即時情緒辨識系統

An online affective cognizition system based on the wireless Brain Computer Interface.

陳博明(Po-Ming Chen)

指導教授 ： 李有璋 ； 劉益宏

中原大學/工學院/機械工程研究所/碩士(2015年)

https://doi.org/10.6840/cycu201500566

摘要

本論文基於腦機介面(Brain-Computer Interface, BCI)搭配無線乾式電極帽，開發出一套即時的情緒辨識系統，能夠準確地反映出不同的情緒，幫助醫師快速診斷病患的情緒狀態。此系統包含三個模式，分別為訓練模式(Training mode)、回饋模式(Feedback mode)、非同步模式(Asynchronous mode)，在訓練模式中，由國際情緒圖片系統(International Affective Picture System, IAPS)所提供的情緒圖片中選出 100 張圖片，分別對 5 位受測者進行情緒誘發，經過每一次的情緒誘發後，受測者皆必須依照自我的感覺填寫自我評估(Self-Assessment Form, SAF)，該評估表中針對受測者填選的 Valence 程度進行主觀情緒的類別標籤，並利用該標籤作為機器學習分類的依據，其中包括 3 種類別，包括 Low-Valence(LV)、High-Valence(HV)和中性情緒(Neutral) ；回饋模式跟訓練模式一樣使用情緒圖片誘發情緒，但回饋模式在每一次的情緒誘發結束之後，會立即的辨別出使用者於情緒誘發時的情緒狀態；非同步模式中，不進行情緒誘發，無時無刻的辨識使用者的情緒，為此系統的最終目的。
由於低錯誤率對於情緒辨識是必要的，因此，有別於傳統的三類別分類方式，本論文使用線性鑑別分析(Linear Discriminant Analysis, LDA)結合邊界層進行三類別的分類，藉由調整邊界層的厚度，求得限制不同錯誤率下的分類率，並使用頻帶功率(Band Power, BP)與共同空間模式(Common Spatial Pattern, CSP)為特徵抽取方式，比較在不同腦區和不同頻帶下的分類率差異，結果顯示low - Gamma 頻帶搭配共同空間模式為特徵抽取，正負類別平均分類率為 77.57%，最高可達 92.43%。最後使用線性鑑別分析搭配一對一共同空間模式進行三類別分類，發現於錯誤率為23.60%下，分類率只有46.35%，反觀本論文所提出之演算法分類結果，在相近的錯誤率下，分類率可達77.57%，具有較好的分類表現。綜合以上結果，對於即時的情緒辨識有著很大的幫助。

關鍵字

腦機介面 ； 情緒辨識 ； 腦波訊號 ； 線性鑑別分析 ； 共同空間模式

並列摘要

This paper developed an online affective recognition system based on Brain-Computer Interface (BCI) with wireless dry electrode cap which can reflect different emotions to assist physicians quickly diagnose the emotional state of patients. The system has three modes including training mode, feedback mode and asynchronous mode. In training mode, we selected 100 emotion pictures from International Affective Picture System (IAPS) to induce the emotion of five subjects. After each emotion induction, subjects have to fill out a self-assessment form (SAF) in accordance with self-feeling, providing the criteria of machine learning classification according to the subjective emotion of the subjects. There are three classes in the classification including High-Valence, Low-Valence and Neutral. In feedback mode, the emotion picture is used to induce emotion as well and the emotional state will be recognized immediately after each emotion induction. Asynchronous mode is the ultimate aim of this study. In asynchronous mode, the emotional state of users will be recognized at all time without emotional induction.
Since a low error rate is necessary for emotion recognition, this paper differs from traditional three-class classification in presenting a Linear Discriminant Analysis (LDA) combined boundary layer as classifier. The classification rates under different error rates were computed by adjusting the thickness of the boundary layer. Band Power (BP) and Common Spatial Pattern (CSP) were used for feature extraction and the classification rate at different lobes and different bands were analyzed. The results showed that when CSP were used for feature extraction, the averaged positive/negative emotion classification rate of each subject can achieve 77.57% and the highest classification rate reached 92.43% in low-Gamma band. Finally, we used the LDA with OVO-CSP for three-class classification. The result showed that the classification rate was 46.35% with 23.60% error rate. In similar error rate, the classification rate of proposed algorithm was 77.57%. Therefore, compared with OVO-CSP, the proposed algorithm has better performance. Summing all the results can provide significant contribution to online affective recognition.

並列關鍵字

BCI ； Emotion Recognition ； EEG ； LDA ； CSP

參考文獻

[26] 黃俊偉, 非同步腦機介面方法比較及即時腦控開關之發展,中原大學機械工程研究所碩士學位論文, 2013.

[29] 高永樺, 情緒腦波分析及腦機介面之開發,中原大學機械工程研究所碩士學位論文, 2013.

[30] 鄭幃騰, 憂鬱症腦波分析,中原大學機械工程研究所碩士學位論文, 2014.

[2] Keltner, D. and Haidt, J., “Social functions of emotions at four levels of analysis” , Cognitive and Emotions, vol. 13, pp. 505-521, 1999.

[4] R. Caton, “The electrical currents of the brain”, The Journal of Nervous and Mental Disease, vol. 2, no. 4, pp. 610, 1875.