以小波轉換鑑別人類情緒腦電波

Abstract

人類情緒的正確鑑別存在著許多的困難，根據每個人所經歷的事物與心情狀態，影響著即使面對相同的事件，所呈現的情緒強度也有所不同。而現今對於人類腦波的研究逐漸盛行，藉由大腦人機介面(Brain computer interface)收集腦電波(Electroencephalogram)訊號，經由訊號分析、特徵擷取以及分類器，來鑑別腦電波訊號的情緒類別。本研究的受測者為六位男性，四位女性。年齡介於20歲至28歲。實驗流程為撥放六種情緒的臉部圖片，分別為高興、驚訝、生氣、厭惡、難過和恐懼，每種情緒有20張圖片，共有120張圖片。使用NeuroScan大腦人機介面，藉由非侵入式的腦電波訊號量測，共有30個通道。紀錄完成後，進行腦電波訊號前處理降低腦電波訊號的雜訊，使得腦電波訊號更接近真實的訊號，接著繪製出大腦空間能量頻譜圖，用以了解腦電波訊號的頻帶能量分布差異。將腦電波訊號進行小波轉換(Wavelet transform)分解訊號，選取能量分布差異較大的θ波為分類波段，接著計算各種的特徵，共有八類特徵，分別為最大值(Max)、最小值(Min)、全距(Range)、標準差(Standard deviation)、絕對中位差(Median absolute deviation)、絕對平均差(Average absolute deviation)、能量(Energy)及特徵向量(Eigenvectors)，將各種特徵投入支持向量機(Support vector machine)進行分類，訓練的方式將隨機抽取出60%的腦電波訊號區段為訓練資料，40%為測試資料，以隨機投入支持向量機作各種情緒的鑑別，得到情緒鑑別從最高到最低的正確辨識率分別為87.50%和62.50%，平均值為76.25%。 研究中發現當使用無效的特徵或是相似的特徵，無法增加情緒的鑑別率，但是若增加有效的特徵，鑑別率會隨之提高，不過也會增加複雜度，經由比較其中較為有效的特徵為全距、標準差、絕對中位差、絕對平均差、能量及特徵向量，可較為明顯增加鑑別的效果。It’s difficult to classify the human emotions correctly. According to the events and mood which have experienced, even at the same event, people present the different emotional intensity. The current studies of prevail in EEG signals are more popular. The brain computer interface recorded EEG signals through feature extraction and classification to identify the type of EEG signals. Four females and six males in the age group of 20-28 years were employed as subjects in our experiment. During the experiment, 120 pictures were shown to the subjects for each 5seconds; 20 pictures per emotion. EEG signals recordings were conducted using the NeuroScan (30 channels). With the pre-process to EEG signals, we took the two trials for both the average, which aimed to reduce noise in EEG signals. And we also mapped the brain space energy spectrum and analyzed and found that EEG frequency θ band was the best one for classification. We used wavelet transform to decompose EEG signals. And then, we calculated all the characteristics features. There are eight kinds of features, including maximum, minimum, range, standard deviation, median absolute deviation, average absolute deviation, energy and eigenvectors. We took these characters into support vector machine for classification. The training data will be randomly selected from 60% of the EEG trials and the testing data will be randomly selected from 40% of the EEG trials. The highest emotional recognition rate we received was 87.50%, but the lowest recognition rate was 62.50%. The average of emotional recognition rate was 76.25%. Since the use of invalid features or similar features, the rate of emotional identification could not be increased. However, if more useful features were used, the rate of emotional identification will be improved. By the comparison of two results, we found that the range, standard deviation, median absolute deviation, average absolute deviation, energy and eigenvectors were useful features.