應用於十六通道腦波即時處理系統之高效能奇異值分解處理器與線上遞迴獨立通道成分分析之系統晶片設計與實作

Abstract

近年來國內外對於腦波的相關研究越來越廣泛，腦波系統已不再止於訊號的截取與分析。腦波系統已經被廣泛應用於醫療診斷、大腦開發、車用電子、遊戲娛樂，腦機介面…等。其中，腦機介面之相關應用更為蓬勃發展。由於腦機介面的成功機會，主要在於是否可得到不受肌肉電波影響以及外界環境干擾的腦電波，而獨立成分分析的主要功能就在於將此類的干擾電波，獨立在與腦波分離的通道上，故經過獨立成分分析的資料處理，即可將乾淨且不受干擾的腦波傳輸到腦機介面系統中，達到更精確的判讀。 獨立成分分析將多重的混合信號分成彼此相互統計獨立的非高斯來源信號，它是一種盲源分離方法，近年來，已成為神經網路、高級統計學和信號處理等研究領域中最常使用的演算法之一。獨立成分分析的目的在於找到一個解混合矩陣W對混合訊號去做運算，進一步求得獨立成份。本研究之晶片設計所採用的是線上遞迴獨立通道成分分析演算法，此演算法提供一種遞迴歸納最小平方近似之線上分析，其特點是一次運算只利用每個通道的一個取樣訊號即可做訓練，其輸出延遲短。比起先前的演算法，不需要做大量的資料暫存，也更具有資料處理的即時性，更適合以晶片的方式來呈現。然而，系統運算過程中必須要大量且重複使用奇異值分解來完成矩陣逆運算。由於奇異值分解為整個系統中運算量最為龐大的處理單元，在高精準度和即時輸出的前提下，奇異值分解的運算時間和效能都會直接影響系統之結果。此系統晶片設計由台積電40奈米製程下線。晶片的核心面積為1.55平方毫米。使用128 Hz的取樣頻率，晶片操作於100 MHz的工作頻率與0.9 V的核心電壓時，功率消耗為9.23毫瓦。 此多通道腦波擷取系統晶片整合了前端晶片控制模組、高效率多通道線上遞迴獨立成份分析處理器、自動肌電雜訊去除模組和後端藍芽傳輸介面。此系統晶片經由前端訊號擷取模組取得混合腦波訊號，並傳送至獨立成份分析處理器將腦波混合訊號分離出獨立成分，並將眨眼雜訊成功分離至某一通道。再藉由自動去除雜訊模組將其消除，最後重建原始的腦波訊號。處理完的數位訊號經由晶片內的藍芽傳輸介面將其打包成四組封包，藉由市售的藍芽模組將這些封包做無線傳輸，並能夠傳輸至臨近的生醫資訊工作站或可攜式行動裝置進行顯像和遠端觀察與診斷。此系統晶片皆以真實生理訊號驗證，並顯示優良的分析結果。此系統晶片提供乾淨的腦波訊號將有利於腦機介面的資料分析，並能夠提升腦機介面的可靠性及可行性。

In recent years, the scope of electroencephalography (EEG) research has expanded. EEG systems are no longer confined to signal extraction and analysis; they have been widely applied to medical diagnosis, brain development, automotive electronics, gaming entertainments, and brain-computer interfaces (BCIs). Specifically, BCI-related applications have been vigorously developed. The key to an effective BCI system is whether EEG signals free of external artifacts signals or influence from muscle noise can be obtained. Independent component analysis (ICA) primarily functions to separate such artifacts from clean EEG signal; therefore, ICA processing results, artifact-free EEG signals, can be transferred to BCI system, thereby attaining precise results. ICA is a technique used for dividing multiple mixed signals into independent non-Gaussian source signals; it is a type of blind source separation (BSS) method. ICA is one of the popular algorithms widely used in research domains such as neural networks, advanced statistics, and signal processing. The objective of ICA is to identify an unmixing matrix W to compute mixed signals and then obtain independent components. In this thesis, a chip was designed based on an online recursive ICA (ORICA) algorithm, which provided an online analysis for recursive least square approximation. This type of ICA features short output delay and the capacity to train data by using a single sample signal extracted from each channel. Compared with other algorithms, ORICA enables real-time data processing, allows results to be presented using a chip, and does not require temporary data storage. However, the singular value decomposition (SVD) processor is widely used to complete the matrix inverse calculation during the system operation in this work. Because the SVD processor spends the most massive amount of computation in the whole system, the computation time and performance of that will directly affect the system result on the premise of the real-time and high accuracy requirements. The System-On-Chip design proposed in this thesis was implemented by using Taiwan Semiconductor Manufacturing Company (TSMC) 40 nm tape-out process. The core size of the chip was 1.55 mm2. The total power consumption was 9.23 mW when a sampling rate of 128 Hz was used, and the chip operated at a frequency of 100 MHz and core voltage of 0.9 V. The multichannel EEG extraction system on chip design developed in this study was integrated with a front-end chip control interface, highly efficient multichannel ORICA training unit, auto de-artifact unit and a back-end Bluetooth interface. This chip extracted mixed EEG signals using the bio-signal acquisition IC, ADS1298, and transmitted these signals to the ICA unit, where the mixed signals were separated into independent components, and eye-blink artifacts were successfully separated to a certain channel. Subsequently, the auto de-artifact removed this artifact and reconstructed the original EEG signal. The processed digital signal was then packaged into 4 sets of packets in the commercial Bluetooth module, HL-MD08R-C2, which were wirelessly transmitted to neighboring biomedical work stations or portable devices for image examination and remote diagnosis. The developed chip was verified using actual physiological signals and yielded satisfactory results. Such system on chip design can provide clean EEG signals that are beneficial for improving EEG data analysis and enhancing the reliability and feasibility of using BCIs.