RAPID: ReAl-time Process ModellIng and Diagnostics: Powering Digital Factories

现代制造业包括高度控制和自动化的过程，精心设计，在严格的规格下，以经济高效和可持续的方式交付产品。为了确保在可变和通常恶劣的条件下的性能，传感器捕获有关过程状态的连续数据流，例如设备和产品。然而，实时分析这些数据的能力提供了目前无法实现的独特优势。学习从传感器数据中校准其运行，监控其健康状态，并对产品结果和维护要求做出准确预测，是未来自主工厂的过程属性。我们已经与希捷（一家主要的硬盘制造商）合作，在他们的工厂设计，开发和实施这种技术，葛兰素史克（一家领先的制药制造商）已经认识到实时过程分析的潜力，以及英伟达（全球GPU提供商）。目标是建立一个生产稳健性水平，以应对造成劳动力数量和供应链连续性不确定性的重大中断和市场波动。这一愿景为响应式制造系统和数字化控制工厂铺平了道路，但也为实现能够无缝分析传感器获得的数据并将其转化为可操作信息的技术铺平了道路。当公司获取大型数据集时，他们处理数据和实时反应的能力受到算法复杂性和数据规模的阻碍。事实上，如果说当前的大流行有什么不同的话，那就是加强了提高制造能力的必要性，以应对需求的突然增加、生产的重新利用，甚至可能是无人驾驶的自主生产。需要在处理能力和制造系统中进行阶跃变化，以便在边缘（即工厂车间）实时分析数据，使其安全，从而通过减少对外部通信和高性能处理资源的依赖来确保更有效的性能。我们建议以一种方法论和安全的方式来完成，对外部因素的依赖最小，从而促使我们研究以实用，成本效益和可持续的方式执行实时分析的方法。RAPID提出了一种双管齐下的方法，通过新颖的“数据草图”算法来降低计算维度，并在GPU技术上使用“透明计算”进行优化，以提供进一步的加速。在与英国希捷和葛兰素史克的详细互动中，我们已经确定了实时分析可以在转变过程和结果中发挥重要作用的制造阶段。特别地，所提议的技术将应用于一个“诊断分析”案例研究，涉及磁盘制造关键计量阶段的光学成像数据，以及两个“预测分析”示例，用于模型学习，以预测硅片的健康状态，并用于改进故障检测、特征提取和化学产品监测。数据草图通过随机采样少量，信息量最大的数据和模型条目，从而大大降低了计算的复杂性，从而可以在很小的精度妥协的情况下快速执行小规模计算。绘制草图在精度和速度上进行权衡，当采样10%左右的数据时，两个数量级的加速是可行的。为了进一步利用这一优势，使用透明计算实现了草图计算，这挑战了传统计算在不需要高精度时进一步加速计算的能力。在工厂环境中使用有噪声的数据和学习过的统计模型进行计算时，对精度的可控降低对于性能改进是谨慎的，但对于噪声鲁棒性也是必不可少的。

Modern manufacturing involves highly controlled and automated processes meticulously designed to deliver products to certain needs within strict specifications and in a cost-efficient and sustainable way. To ensure performance in variable and often harsh conditions, sensors capture continuous data streams about the state of the process, e.g. equipment, and the product. The ability to analyse this data in real time, however, offers unique advantages that are currently out of reach. Learning to calibrate its operation from sensor data, monitor its health status and make accurate forecasts on product outcomes and maintenance requirements, are process attributes of future autonomous factories. We have teamed up with Seagate, a major manufacturer of hard drives, to design, develop and implement such a technology in their factory, GSK, a leading pharmaceutical manufacturer, who have recognised the potential of real-time process analytics and NVIDIA, the global GPU provider. The goal is to establish a level of production robustness against major disruptions and market volatility that create uncertainties on workforce numbers and supply chain continuity. This vision paves the way for responsive manufacturing systems and digitally controlled factories but to materialise technology that can seamlessly analyse sensor-obtained data and translate it to actionable information. Whilst companies capture large datasets, their ability to process them and react in real-time, is hindered by the algorithms' complexity and scale of the data. Indeed, if anything, the current pandemic has reinforced the need to enhance manufacturing capability to cope with sudden increases in demand, production repurposing, and possibly even unmanned, autonomous production. A step change is needed in the processing capability and manufacturing systems where the data can be analysed in real time at the edge, i.e. on the factory floor, making it secure and thus ensuring more effective performance by being less reliant on external communications and high-performance processing resources. We propose that this be done in a methodological and secure way with minimal dependencies on external factors, thus prompting us to investigate ways of performing real-time analytics in a practical, cost-effective and sustainable manner. RAPID proposes a two-pronged approach to reduce the computational dimensionality through novel 'data sketching' algorithms and optimisation using 'transprecision computing' on GPU technology to provide further acceleration. In detailed interactions with Seagate and GSK both based in the UK, we have identified manufacturing stages where real-time analytics can play a major part in transforming processes and outcomes. In particular, the proposed technology will be applied to a 'diagnostic analytics' case study involving optical imaging data for a critical metrology stage in disk manufacture and two 'predictive analytics' examples for model learning to predict the health state of silicon wafers and for improved fault detection, feature extraction and monitoring of chemical products. Data sketching dramatically reduces the complexity of computations by randomly sampling few, the most informative, data and model entries leading to small-scale computations that can be performed very quickly with a small compromise on precision. Sketching trades off precision and speed, and if done optimally a two-order of magnitude speedup is feasible, when sampling around 10% of the data. To exploit this advantage further, the sketched computations are implemented using transprecision computing that challenges traditional computing to further accelerate computations when high precision is not required. In computing with noisy data and learned statistical models in factory environments, a controllable reduction in precision is prudent for performance improvement but also essential for noise robustness.