Signal Processing of Electromyography with Bayesian Framework for Next Generation Motion Integration Human Machine Interface

人机界面(HMI)将很快嵌入我们社会的方方面面。一种类型的HMI用于捕捉、监控、辅助和增强人体运动。它现在被发现在大量不同的领域，包括但不限于制造业、医疗保健、健身跟踪、办公室人体工程学和娱乐业。这里，这种类型的HMI被称为运动集成HMI(MiHMI)。下一代miHMI需要比以往任何时候都更快、更可靠地感知人体运动意图。然而，为了实现这一目标，除了目前仅使用的机械传感器外，还必须采用其他传感器模式。最近，肌电(EMG)--肌肉活动的电表现--就是为了达到这个目的而被研究的。从理论上讲，表面肌电信号是一种比机械传感器更好的快速、可靠的人体运动意图估计来源，因为1)表面肌电信号总是比相应的机械信号出现得更早；2)表面肌电信号总是可测量的，而在某些情况下机械信号是难以测量或不可能测量的。这一局限性阻碍了表面肌电信号作为下一代miHMI快速准确的人体运动意图估计器的全部潜力。虽然在目前的处理框架下有提高表面肌电信号信噪比的方法，但改善幅度很小，结果与标准机械传感器的信噪比相去甚远。因此，迫切需要大幅提高表面肌电信号处理算法的信噪比。为了解决这个问题，我建议采用一种截然不同的范式来处理贝叶斯框架下的表面肌电信号。*拟议的研究计划将对贝叶斯框架在表面肌电信号处理中的理论和应用进行系统的研究。该程序将从一个明确定义的表面肌电信号生成模型开始，申请人已经开发出该模型并成功地应用于灵巧上肢假体(一种特定类型的miHMI)的控制。该模型将扩展到包括更复杂的运动任务。然后，该模型的各种参数的概率属性将被明确地合并到基于贝叶斯的算法中，从而产生用于从表面肌电信号估计运动意图的特定于模型和任务的贝叶斯算法。这些算法将在从多通道表面肌电信号估计运动意图的背景下进行检查、测试和评估。算法的性能将通过估计精度、在线计算速度和对非平稳因素的稳健性来衡量。为了确保算法的实用性，将特别关注贝叶斯算法和用户的长期自适应特性。

Human machine interface (HMI) will soon be embedded in every aspect of our society. One type of HMIs is used to capture, monitor, assist, and augment human motion. It is are now found in a vast array of different fields including, but not limited to, manufacturing, healthcare, fitness tracking, office ergonomics, and entertainment. Here, this type of HMI is called motion integration HMI (miHMI). Next generation miHMIs need to sense human motion intentions faster and more reliably than ever before. To achieve this goal, however, alternative sensor modalities beyond mechanical sensors, which are exclusively used currently, must be incorporated. Recently, electromyogram (EMG), the electric manifestation of muscle activities, has been investigated for this purpose. In theory, surface EMG (sEMG) is a better source for fast and reliable human motion intention estimation than mechanical sensors because 1) it always appears earlier than the corresponding mechanical signals; 2) it is always measureable, while in some cases mechanical signals are difficult or impossible to measure.***However, current sEMG processing framework can only achieve an extremely low signal-to-noise ratio (SNR), orders of magnitude lower than that of a standard mechanical sensor. This limitation prevents sEMG from reaching its full potential as a fast and accurate human motion intention estimator for next generation miHMIs. Although there are ways to enhance sEMG's SNR under the current processing framework, the improvement is marginal, and the result is nowhere near the SNR of standard mechanical sensors. As such, there is an imperative need for drastic improvements to the SNR of sEMG processing algorithms. To address this problem, I propose to take a radically different paradigm for processing sEMG with the Bayesian framework. ***The proposed research program will undertake a systematic investigation in the theory and application of the Bayesian framework in sEMG processing. The program will start from a well-defined generation model for sEMG, which the applicant had developed and successfully applied to the control of dexterous upper limb prosthesis (a specific type of miHMI). The model will be extended to incorporate more complex motor tasks. Then, probabilistic properties of various parameters of this model will be explicitly incorporated within Bayesian-based algorithms, resulting in model and tasks specific Bayesian algorithms for estimating motion intentions from sEMG. These algorithms will be examined, tested, and evaluated, in the context of estimating motor intentions from multi-channel sEMG. The performance of the algorithms will be gauged by estimation accuracy, online computation speed, and robustness against nonstationary factors. To ensure the practical applicability of the algorithm, special focus will be on the long term adaptive characteristic, both of the Bayesian algorithms and from the user.* **