SBIR Phase I: A Fully Autonomous Prognostic Digital Twin for Smart Manufacturing

这个小企业创新研究（SBIR）第一阶段项目的更广泛/商业影响旨在帮助行业减少计划预防性维护的停机时间。拥有高价值资产（如制造设施、发动机、卫星、反应堆等）的行业，由于缺乏对生产力优化的可用见解，通常会产生大量费用。该项目产生的预测技术和发展将具有普遍适用性，并能够在不同的市场中使用说明性预测（何时和修理什么）。此外，项目中开发的用于在有限数据上训练深度学习系统的方法将在机器学习（ML）社区中具有广泛的应用。项目经常受到数据获取和可用性的限制。本项目开发的方法可以应用于小型医疗数据或金融数据集，因为它们完全是在时间序列变量和动态上定义的。SBIR第一阶段项目有两个主要目标。首先，开发一种技术，能够完全自主地从原始传感器数据中提取有意义的特征集。本项目开发的自主特征选择程序将利用强大的控制理论结果与现代机器学习工具的结合来发现非明显的线性和非线性特征。该解决方案将提供一个物理信息的架构，允许用户将可用的物理知识与数据中出现的知识结合起来，配置一个健壮、灵活和自主的特征提取机制。其次，该团队将构建一个鲁棒的、多模态的传感器模拟器，以解决数据不足的问题，从而训练机器学习组件。这个机会是为了应对制造业数据的有限可用性，特别是操作系统中的时间序列传感器数据。传感器模拟器将通过现代基于ml的生成工具的组合形成，以一种利用其已证明的有效性的方式，同时能够处理高维信号和小型训练数据集。该奖项反映了美国国家科学基金会的法定使命，并通过使用基金会的知识价值和更广泛的影响审查标准进行评估，被认为值得支持。

The broader/commercial impact of this Small Business Innovation Research (SBIR) Phase I project seeks to assist industries reduce their downtime for scheduled, preventative maintenance. Industries with high-value assets like manufacturing facilities, engines, satellites, reactors, etc., often incur significant expense due to a lack of usable insights into productivity optimization. The forecasting technology and the developments stemming from this project will have general applicability and enable the use of prescriptive prognostics (when and what to repair) in diverse markets. Additionally, the methods developed in the project for training deep learning systems on limited data would have broad application within the machine learning (ML) community. Frequently, projects are limited by access to and availability of data. The methods developed in this project could be applied to small sets of medical data or financial data, as they are entirely defined on time series variables and dynamics. This SBIR Phase I project has two main goals. First, to develop a technology that will enable full autonomy in the extraction of meaningful feature sets from raw sensor data. An autonomous feature selection procedure developed in this project will exploit the combination of powerful control-theoretic results with modern ML tools to discover non-obvious linear and nonlinear features. This solution will provide a physics-informed architecture, allowing users to incorporate available physics knowledge with that emerging from the data, configuring a robust, flexible, and autonomous feature extraction mechanism. Second, the team will construct a robust, multi-modal, sensor emulator to address data insufficiency in order to train the ML components. This opportunity is in response to the limited availability of data in manufacturing sector, especially time-series sensor data in operational systems. The sensor emulator will be formed via combinations of modern ML-based generative tools in a manner that exploits their proven effectiveness while being able to work with high-dimensional signals and small training datasets.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.