具有自動雜訊去除機制之32多通道腦波擷取系統晶片設計

Abstract

在此篇論文中，我們提出了一個可具有自動雜訊去除機制之即時32多通道腦波擷取系統晶片設計。腦波訊號是非常微弱的電訊號，故在擷取腦波的過程，經常受到非腦波成份的電訊號干擾。利用獨立通道成分分析演算法從一段時間的腦波訊號中萃取出包含於腦波訊號中的雜訊。經過獨立通道成分分析分離出各個來源後，我們可以去除含有雜訊的通道，來重建出不含非腦內成分的腦波訊號。 近來，腦機介面(BCI)蓬勃發展。這類的介面使人可以經過大腦直接控制機器。為了增加此類介面的發展性，以及其運作的穩定性、正確性，即時的擷取到不含非腦內成分的腦電訊號是個重要的課題。 線上遞迴式獨立成分分析這樣的演算法，提供了我們一個即時雜訊萃取的機制，能在每次腦波訊號由前端擷取器完成取樣後，隨即完成獨立通道成分分析的演算，因此非腦內成分的雜訊能夠及時的被萃取出來。因為這樣即時萃取雜訊的特性, 再經由 (Blind source separation- Canonical Correlation Analysis, BSS-CCA) 進行肌電及眼動雜訊的去除, 即可將乾淨且不受干擾的腦波傳輸到腦機介面中進行處理, 達到更精確的判讀. 干擾腦波擷取的雜訊大致上可分為兩類：人類自發性的電波干擾，以及外在環境的電雜訊干擾。眾所周知，經由骨骼肌產生並由皮膚擷取的肌電雜訊影響α，β，甚至θ頻段，其頻域之寬從0到> 200赫茲皆或多或少有不同的頻譜分量。然而，許多自動偵測與去除EMG的方法多著重於影響人腦波約15到20Hz的雜訊干擾。其中額和顳肌肉這兩種肌肉的肌電為最常見的來源。因此在此設計中，我們針對這樣的干擾進行去除。現今的演算法中，雖然具備了自動去除雜訊的機制，但應用於人機介面中無法使我們享有完全即時性、可攜性。在此設計中，我們利用了前人提出的演算法，對演算的流程進行修改，使其能進行即時的肌電雜訊去除。 此演算法設計於TSMC 40 nm COMS 製程，並實現於FPGA上。每個腦波訊號能於取樣後的0.27秒內得到不含肌電雜訊的腦波訊號，在此論文中我們亦提供了評估此系統效能的方式，針對在消除肌電與眼動雜訊的效能上，與原始腦波的平均相關度為0.9341 和 0.8927，另外我們亦使用真實的腦波進行處理，結果顯示，經過此系統的處理後，雜訊確實的被去除。

This thesis presents a system-on-chip design of a 32-channel EEG acquisition system with an automatic artifacts rejection. EEG signals are among the feeblest physiological electrical signals we can capture from the human body and are easily contaminated with artifacts caused by noncerebral electrical activities. Previously, ICA was used to extract artifacts from a period of EEG data, and after processing ICA, An automatic artifacts detection and elimination algorithm is performed to remove the artifacts from EEG signals and free EEG signals can be efficiently reconstructed. Nowadays, brain-computer interfaces (BCIs) are being developed to control machines through EEG analysis directly. To enhance the accuracy, and feasibility, reliability of BCIs, EEG signals used for BCI applications should be acquired in real-time from the human scalp without artifacts. The real-time requirement is achieved by adopting online recursive ICA (ORICA) which can help extract artifacts in real-time as it can immediately produce ICA results right after each EEG sample acquisition. Then through BSS-CCA (Blind source separation- Canonical Correlation Analysis) we can remove the artifacts and get a clean and undisturbed EEG signal ready to be processed by brain-computer interfaces with better and more accurate interpretations. There are two kinds of artifacts. The first type originates from the inside of the human body and is referred as biological artifacts, typically muscle artifact known as EMG artifact. The second one is due to the surrounding environment, from outside of the human body, and is referred as environment artifact. It is known that EMG of skeletal muscles recorded from the skin has a broad frequency distribution from 0 to >200 Hz with several more or less distinct spectral components. EMG activity affects alpha, beta, and even delta frequency bands. However, most automated methods for EMG artifact detection and elimination make the assumption that EMG recorded from the scalp has a broad spectral distribution that begins at 15–20 Hz. Therefore, both artifacts from frontalis and temporalis muscles are the ones we focused on. These two muscles are the most common sources of EMG artifacts over frontal and central head regions when recording EEG signals. They must be removed to avoid any misleading in the operation of BCIs. The system is designed to include these algorithms using TSMC 40nm CMOS technology and implemented on FPGA. With the proposed multi-channel processing flow, multi-channel artifact free EEG signals ca be acquired in 0.27 s after input sample. The performance of eye blink artifact and muscle artifact elimination is evaluated through the computation of correlation coefficients between original artifact free EEG signals and processed artifact free EEG signal which are 0.9341 and 0.8927 averagely. The processed results with real EEG signals are also shown to remove muscle artifacts successfully.