| Brain Computer Interface(BCI) is a fast-growing emergent technology, in which researchers aim to build a direct channel between the human brain and the computer. The efficiency of a BCI should encompass three stages: the first is recording of the cerebral signal; the second is extraction of mental task related features or information from the recorded signal; and the third is translation of the extracted information into a control command. Among these, the feature extraction is the most crucial step. In this paper, we study mainly research of EEG feature extraction method which based on cognitive status recognition.Firstly this paper designed a cognitive experiment which contains 11 kinds of cognitive tasks, collected a number of subjects of EEG data. Due to the property of variety, high resolution, massive data, the data set of has reached to the international level.Secondly, we proposed a CSP feature extraction method which based on time-frequency power features. Comparing with the traditional CSP methods, the proposed method not only retained the time domain variance features, but also induced the frequency band power features. The method obtained high recognition rate in five task cognitive status data sets. Aiming at recognizing the cognitive states quickly and accurately in brain computer interface(BCI), this paper also presents an Electroencephalography(EEG) feature extraction method, which based on common spatial pattern(CSP) and nonnegative matrix factorization(NMF) method. For the problem of Multi-channel, massive data and data redundancy. First of all, CSP method was applied to project the original EEG signals to the common space which has the biggest difference between two kinds of motor imagery tasks; The second, extract the time-frequency energy as global features, which can localize the signal energy in both the time and frequency directions. The NMF algorithm was used to extract the local features. Then fusing the global and local features. The F-score method was used to evaluation criterion of feature selection; At last, support vector machine(SVM) was used for classification. The performance of the proposed method is compared against two existing algorithms, s CSP and KLCSP, using the publicly available BCI competition III dataset IVa. The results show that the proposed method significantly outperforms both the s CSP and the KLCSP algorithms in terms of classification accuracy, which achieved 87.18%. This proposed method in this paper provides a new framework for feature extraction problems of EEG signals.Finally, this paper based on the background of human cognitive state classification and recognition, investigates a hybrid feature evaluation and selection method based on CSP and F-score. Since F-score is easy and fast to calculate, and the F-score based method can quickly select more effective features from high dimensional data, depending on the importance of each data pattern. Face the conventional F-score problem of threshold definition, an automatic method is proposed on the integral tendency of F-score value evolution for data patterns. An optimization method for removing the redundant information is also proposed based on the principle of maximum relevance vs. minimum redundancy of big data information. Combining these methods, a hybrid CSP/F-score common spatial pattern selection framework is constructed, which appears to be automatic and self-adaptive.The proposed method was tested in a cognitive state analysis problem, where involved five mental tasks as baseline, complex problem solving, geometric figure rotation, mental letter composing and visual counting. The recognition accuracy achieved is above 92%. We have also gotten satisfactory results in the data sets of multitasking cognitive state which build in this paper. This approach may provide a new and powerful tool for EEG feature extraction problem. |
PDF Full Text Request