SBIR Phase I: Software with Breakthrough Composite Distance Method for Zero Defects in Advanced Manufacturing

该SBIR第一阶段项目旨在开发基于机器学习的新型软件，用于汽车和医疗器械行业的早期预警和消除潜在的现场故障。软件核心的复合距离技术对于早期识别这些行业的故障单元至关重要，这些行业的现场故障有受伤或死亡的风险。在汽车行业，涉及伤亡的现场事故数量激增。在医疗器械行业，在过去的四年里，五分之四的I级回忆--即那些导致重伤或死亡的事故，是由于复杂的电子故障造成的。随着进入汽车和医疗设备的电子产品在产品中占据越来越大的份额，同时制造过程变得更加复杂，这使得在部件/设备发货之前检测缺陷变得困难。这个项目中正在开发的软件将检测由于设备中许多细微缺陷的综合影响而产生的缺陷。这些缺陷不是使用目前在制造中进行的标准测试来检测的。该项目符合国家科学基金S的指导，支持具有实质性社会效益的先进制造业的创新和转型技术。该项目完成后，该技术将使制造商能够检测并消除现场故障概率较高的设备，从而保护司机和患者的生命安全。除了对社会的这些好处外，这项技术的商业化还将有助于税收，并为数十名工程师和管理人员创造就业机会。由于许多影响/变量对设备的综合影响而导致的现场故障非常难以检测--这些设备在制造过程中通过了所有规格(否则就不会发货)。在汽车和医疗设备市场，现场故障可能是灾难性的，可能导致生命和肢体的损失。在这个项目中，我们开发了突破性的技术来检测和标记预计将在下游发生故障的单元，同时通过所有当前的规范和控制限制。这一独特算法的有效性在一场由主要分析参与者参与的现场客户评估比赛中脱颖而出。在这项评估中，复合距离预测精度最高，产量损失成本最低。这种方法有两个主要步骤？变量约简和复合距离计算，而我们使用专有方法在这两个步骤之间迭代，以得出用于计算复合距离(CD)的重要变量。该参数CD是在制造过程中为每个单元计算的，反映了所有重要变量的相互作用，并提供了异常行为的测量，并允许识别现场故障可能性高的特立独行的不符合模式的部件。知识产权，因此新颖性，在于识别重要变量的方式，以及迭代删除只会增加噪音的不重要变量的方式。本项目的目标是制作一个实时异常检测的演示软件，具有低延迟的大数据能力。

This SBIR Phase I project aims to develop novel machine learning-based software for early warning and elimination of potential field failures in automotive and medical device industries. The Composite Distance technique at the heart of the software is crucial for early identification of failing units for these industries where field failures carry the risk of injury or death. In the automotive industry, there has been a surge in the number of field incidents involving injury or death. In the medical device industry, in the last four years, four out of five class I recalls -i.e. those leading to severe injury or death, are due to the failures from complex electronics. As the electronics going into cars and medical devices take up an increasing share of the product while simultaneously manufacturing processes become more complex, this is making it difficult to detect defects before units/devices are shipped. The software being developed in this project will detect defects resulting from the combined effects of many subtle flaws in the device. These defects are not detected using standard testing currently done in manufacturing. This project is in line with the National Science Foundation?s direction to support innovative and transformational technology for advanced manufacturing that has substantial benefits to society. On completion of this project, the technology will enable manufacturers to detect and eliminate devices that have a high probability of failing in the field, thereby protecting the lives of drivers and patients. Aside from these benefits to the society, commercialization of this technology will contribute to tax revenue and create jobs for dozens of engineers and managers. Field failures resulting from combined effects of many influences / variables on the unit are extremely difficult to detect - these units pass all specifications during manufacturing (or else they would not have been shipped). In the automotive and medical device markets field failure can be catastrophic and can result in loss of life and limb. In this project we develop breakthrough technology to detect and flag units that are predicted to fail downstream while passing all current specifications and control limits. The effectiveness of this unique algorithm stood out in a competition involving major analytics players in an onsite client evaluation. In this evaluation the Composite Distance produced the highest predictive accuracy and lowest cost due to yield loss. The method has two major steps ? variable reduction and Composite Distance computation, while we use proprietary methods to iterate between these two steps to arrive at the key variables of importance that are used to calculate the Composite Distance (CD). This parameter CD, computed for each unit during manufacturing reflects the interaction of all variables of importance and provides a measure of anomalous behavior and allows to identify maverick out-of-pattern parts with high likelihood of field failure. The intellectual property, and therefore the novelty, lies in the way important variables are identified and unimportant variables, that only serve to add noise, are removed iteratively. The goal of this project is to produce a demonstration software for real-time anomaly detection, with low latency big data capability.