Using non-negative matrix fact factorization to extract attention-related EEG features

Abstract

The fundamental of non-negative matrix factorization algorithm was introduced. It is used to extract EEG power spectrum feature. Artificial neural network is employed as classifier. Three level attention mental tasks are designed to test the method. Ten subjects attended the experiment. The classification accuracies indicate that the NMF technique is a powerful feature extractor in high-dimensional feature space. The average classification accuracy of ten subjects achieves 88%, it is higher obviously than that of principal component analysis and direct method.

介紹了非負矩陣分解算法(NMF)的基本原理,給出一種利用NMF進行腦電能量譜特征提取的方法。設計試驗對10個被試在三種不同注意任務中的腦電信號進行特征提取,并采用人工神經網絡作為分類器進行分類測試。結果表明,NMF算法在高維特征空間具有較強的特征選擇能力,其分類正確率明顯高于主分量分析(PCA)方法和直接法,三種意識任務的分類正確率分別達到84.5、88%和86.5。