The Finger Flexion Related Feature Extraction Method Based on Wavelet Time-Frequency Analysis in ECoG Signals

In the brain-computer interface system, the feature extraction of brain signals is a crucial procedure. Especially in the multi-channel brain signals such as Electroencephalogram (EEG), Electrocorticography (ECoG), the channel which has the most correlation with the goal human activity and intention is the priority concern. However, because of the complicated extraction to the feature of the human fine part movements, most of the previous studies are aiming at the imaginary or real activity of large body parts, and their features are usually used in classification tasks. Thus, in order to extract the feature which has a higher linear correlation with fine body part such as fingers, this paper proposes a method combining wavelet time-frequency analysis and principal component analysis (PCA) to extract finger flexion related feature. In the first step, the multi-channel signals will be pre-processed. Then the time-frequency spectrum of each channel's signal is calculated by continuous wavelet transform. After that the spectrum is optimized, and the first wavelet time-frequency spectrum principal component (Wtspc) is extracted by PCA. At last, the Wtspc, which has the highest correlation to the corresponding finger flexion, is chosen as the final feature. The experiment results indicate that the Wtspc feature which extracted by our method has a higher correlation than original signals and typical time-domain features in the previous studies. Particularly, in the local finger flexion period, the Wtspc feature highly demonstrates a linear correlation with corresponding finger flexion.