Application of Probabilistic Sensor Networks to Intent Detection in Prosthetics and Other Wearable Technology

项目背景和潜在影响在英国，有超过6万名截肢或先天性肢体缺陷患者在专业康复服务中心接受治疗，假肢技术是一个重要的行业，有可能影响许多人的生活。然而，很明显，在假肢控制方面还有很大的改进空间。典型的健康人不会决定按顺序单独激活每一块肌肉来执行一个动作——他们只是“想”出来，肢体就会自动服从。相比之下，操作假体可能具有挑战性，需要大量的实践来执行任何类型的复杂操作，这是假体设计的限制因素，并可能导致设备放弃。该项目旨在开发技术，使假肢从“手动”控制转向计算机辅助系统，更让人想起健康的生物肢体控制。项目目标和目标该研究的目标是设计、开发和测试一个系统，用于概率地结合来自多个传感器环境（如肌电图、运动跟踪、运动学传感、智能家居和基于智能手机的传感）的读数，以便在操作可穿戴设备（如假肢）时产生用户意图的测量，然后可用于决定设备应如何响应。目标是这将是一个“插入/退出”网络，利用任何给定时间可用的任何传感器，专注于低成本，不引人注目的传感器。这一概念将在一系列特定的、可重复的操作的“测试场景”中得到证明，并扩展为一组可以应用于广泛情况的一般原则。我们进行了广泛的文献综述，评估了所有现有的意图感知研究，并确定了所有可用于网络的可行传感器方法的优点。对概率传感器网络技术进行了回顾，并对我第四年项目的结果进行了推进和扩展，以产生一套概率意图感知系统的理论模型。利用这些模型，已经开发出一种方法，为机器学习系统提供“安全性检查”，以识别不适当的方法（在UEMCON 2019和EMBC 2020上发表并介绍）。这些模型和方法正被用于开发意图感知算法，并将应用于坐下、开始爬楼梯和拿起杯子等测试场景，这些场景最初将被人工模拟。一旦系统被证明可以在模拟中工作，实验数据将从实际的传感器网络中收集，以在现实世界中进行测试。其中一些数据将从健康的志愿者那里收集，布拉奇福德（一家英国领先的假肢公司）可能会提供来自患者的额外数据，这些数据可以用来验证系统，并测试它的成功如何因用户而异。随着该概念在多个测试场景中被证明有效，该过程将被形式化为一套通用技术，可通过输入各种数据参数用于广泛的应用，采用机器学习技术来训练系统将传感器数据与意图相关联，无论情况如何。该框架旨在使工业应用更加简单，并将与布拉奇福德公司合作，以确保这些原则在以产品为中心的环境中是可行的。与EPSRC的战略和研究领域保持一致该项目属于EPSRC工程和医疗保健技术研究领域，不仅在假肢控制和患者康复领域具有潜在的应用，而且在可穿戴技术和人机接口等更广泛的领域也有潜在的应用。

Project Context and Potential ImpactWith over 60,000 patients with an amputation or congenital limb deficiency attending specialist rehabilitation service centres in the UK, prosthetic technology is a significant industry with the potential to impact a great many lives. However, it is clear that there is much room for improvement in prosthetic control. The typical, healthy human does not decide to activate each muscle individually in a sequence to perform an action - they simply "will" it, and the limb obeys automatically. In contrast, operating a prosthetic can be challenging, requiring considerable practice to perform any kind of complex operation, which is a limiting factor in prosthetic design and can contribute towards device abandonment. This project will aim to develop techniques to move away from "manual" prosthetic control to a computer-assisted system more reminiscent of healthy biological limb control. Project Aims and ObjectivesThe objective of the research will be to design, develop and test a system for probabilistically combining readings from multiple sensor environments (such as EMG, motion-tracking, kinematic sensing, smart home and smart phone-based sensing) in order to produce a measure of user intent when operating a wearable device such as a prosthetic limb, which can then be used to make a decision as to how the device should respond. The goal is that this will be a "drop in/drop out" network, taking advantage of whatever sensors are available at any given time, focusing on low-cost, unobtrusive sensors. This concept will be proved over a number of "test scenarios" for specific, repeatable actions, and expanded into a set of general principles which can be applied over a wide range of situations. Proposed Research MethodologyAn extensive literature review has been carried out, assessing all existing intent-sensing research and determining the merit of all viable sensor methods that could be used in the network. Probabilistic sensor network techniques have been reviewed, carrying forward and expanding upon the results of my fourth year project to produce a set of theoretical models of probabilistic intent-sensing systems. Using the models, a method has been developed to provide a "sanity-check" for machine learning systems to identify inappropriate methods (published and presented at UEMCON 2019 and EMBC 2020). These models and methods are being used to develop the intent-sensing algorithm and will be applied to test scenarios such as sitting down, beginning climbing stairs, and picking up a cup, which will initially be artificially simulated. Once the system is proven to work with the simulation, experimental data will be gathered from practical sensor networks to test it in the real world. Some of this will be gathered from healthy volunteers, and additional data may be provided by Blatchford (a leading UK prosthetics company) from patients, which can be used to verify the system and test how its success may vary from user to user. With the concept proven to work in multiple test scenarios, the procedure will be formalised into a general set of techniques which can be used for a wide range of applications by inputting various data parameters, employing machine learning techniques to train the system to associate sensor data with intent, regardless of the situation. This framework will be designed to make industrial application straightforward, and cooperation will take place with Blatchford to ensure the principles are viable in a product-focused environment.Alignment to EPSRC's Strategies and Research AreasThis project falls within the EPSRC Engineering and Healthcare Technologies research areas, with potential applications not only in the fields of prosthetic control and patient rehabilitation, but also in the wider fields of wearable technology and human-machine interfacing as a whole.