Collaborative Research: Micro-Electro-Mechanical Neural Integrated Sensing and Computing Units for Wearable Device Applications

随着可穿戴设备在消费市场获得吸引力，对健康和行为进行不引人注目的持续监测直接转化为改善健康和生活质量。这些平台提供了新的机会来检测疾病的早期发作、评估人类表现或提高生产力，以及许多其他潜在的应用。然而，这些设备面临的主要挑战是电池续航时间。由于严格的空间要求，这些设备中的电池很小，通过执行复杂的算法(例如机器学习)和大量的无线通信可以迅速耗尽电池。反过来，这又迫使用户更频繁地向他们收费，并阻碍了这些设备的广泛采用。为了克服这一挑战，拟议项目的目标是突出微电子机械系统(MEMS)设备作为混合传感和计算元件的计算潜力，使可穿戴设备能够高效地执行复杂的算法，同时保持电池电量。该项目提出了一种新的、高度智能的计算单元技术，可以由永久性电池供电，并可以整合到许多医疗应用中，从而具有巨大的潜力来影响美国工业。将包括内布拉斯加州-林肯大学、达拉斯德克萨斯大学和德克萨斯农工大学在内的三所机构聚集在一起。这一项目的成果还将用于所有三个机构正在教授的各种课程。它还将用于2020年开始的NanoBridge夏令营，通过MEMS和纳米工程方面的教育活动，提高来自代表性不足群体的高中生的工程兴趣。该项目旨在开发一种用于可穿戴设备的超强计算单元，用于在本地执行机器学习算法。算法将编码在MEMS的机械响应中，同时捕捉感兴趣的测量，如加速度。配备机器学习算法的可穿戴设备在拯救生命方面具有巨大的潜力，例如，通过自动检测跌倒。然而，由于严格的空间要求，这类设备中的电池很小，并且在很大程度上被多个MEMS传感器读出电路、无线通信和微处理器迅速耗尽。这直接导致了不遵守规则，因为用户必须频繁地为他们的设备充电，并且可能会遇到由于本地计算能力有限而必须使用的较不准确的算法导致的错误警报。为了克服这些挑战，提出了一种新的方法，将部分计算转移到传感物理层。这种方法基于这样一个事实，即MEMS设备的传感元件只需要很少的功率，其与其他传感元件耦合的机械响应可以进行调整，以根据自己的测量自然执行机器学习算法。因此，多个感测元件的响应将共同编码高级信息，而不是产生需要被放大、调节并从模拟到数字转换以由微处理器读取和处理的行测量信号。这种方法将使可穿戴设备能够在本地执行先进的算法，同时消耗的电力比目前最先进的技术低两个数量级。这一奖项反映了NSF的法定使命，并通过使用基金会的智力优势和更广泛的影响审查标准进行评估，被认为值得支持。

As wearable devices gain traction in the consumer market, unobtrusive and continuous monitoring of health and behavior directly translate to improving wellness and quality of life. These platforms provide new opportunities to detect the early onset of a disease, assess human performance, or enhance productivity, among many other potential applications. The principal challenge with these devices, however, is their battery life. Due to stringent space requirements, the batteries within such devices are small and can be quickly drained by performing sophisticated algorithms (e.g. machine learning) and heavy wireless communications. This, in turn, forces users to charge them more frequently and discourages widespread adoption of these devices. To overcome this challenge, the goal of the proposed project is to highlight the computational potential of micro-electro-mechanical-systems (MEMS) devices as hybrid sensing and computing elements to enable wearable devices to efficiently perform sophisticated algorithms while preserving their battery power. This project has tremendous potential to impact US industry by bringing forward a new, highly-intelligent computing unit technology that can be powered by a permanent battery and can be incorporated into many medical applications. Bringing together three institutions including the University Nebraska-Lincoln, the University of Texas at Dallas, and Texas A&M University. The results of this project will also be adopted into various courses being taught at all three institutions. It will also be used in a NanoBridge summer camp beginning in 2020 to promote engineering interest among high school students from underrepresented groups through educational activities in MEMS and nanoengineering.This project aims to develop an ultra-power computing unit for wearable devices to locally perform machine-learning algorithms. The algorithms will be coded in the mechanical responses of MEMS that also simultaneously capture the measurement of interest, such as acceleration. Wearable devices equipped with machine learning algorithms hold great potential for saving lives, for example, by automatically detecting falls. However, due to stringent space requirements, the batteries within such devices are small and are quickly drained, for the most part, by multiple MEMS sensors read-out circuity, wireless communication, and microprocessors. This contributes directly to nonadherence as users must charge their devices frequently and may have trouble with false alarms caused by the less accurate algorithms that must be used due to limited local computing power. To overcome these challenges, a novel approach is proposed that moves some of the computing to the sensing physical layer. This approach builds on the fact that the sensing element of a MEMS device requires very little power, and its mechanical response coupled with other sensing elements can be tuned to naturally perform machine learning algorithms from their own measurements. Thus, rather than producing row measurement signals that need to be amplified, conditioned, and converted from analog to digital to be read and processed by a microprocessor, the response of the multiple sensing elements will collectively encode high-level information. This approach will enable wearable devices to locally perform advanced algorithms while consuming two orders of magnitude less power than present state-of-the-art technology.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.