CPS: Small: Mitigating Uncertainties in Computer Numerical Control (CNC) as a Cloud Service using Data-Driven Transfer Learning

计算机数字控制(CNC)是现代制造机床的重要特征。它提供基于一组编程指令的自动控制，这些指令传统上运行在物理上与机器相连的本地计算机上。这项工作展望了未来，通过安装在云计算机上的CNC，通过互联网远程控制制造机器。在与传统数控相比的几个优点中，基于云的数控有望以低成本显著提高制造机器的速度和精度。然而，基于云的数控的一个主要挑战是，有点像视频流，它主要使用预先计算的命令来控制制造机器，这些命令必须进行缓冲，以缓解互联网传输延迟。因此，基于云的数控系统很容易出现异常，这些异常是由于有关受控机器实际行为的信息传输延迟而导致的。该奖项支持对从过去经验收集的数据预测即将到来的异常情况的方法进行科学调查，并使用预测来避免由于反馈不足而导致的不正确控制行动。从传统数控向基于云的数控过渡，美国将在经济上受益，因为到目前为止，美国是基于云的服务的市场领先者。该项目还将包括接触美国公司，开发教育课程以增加美国制造业和数据分析方面的人才库，以及为底特律地区的中学生开展活动，以激励他们追求工程领域的职业生涯。该项目的目标是减少与使用数据驱动的迁移学习从云中实时控制制造机器相关的不确定性。所获得的知识将为机床提供可靠的基于云的数控系统，从而以低成本提高机床的性能。在基于云的数控系统中，先进的前馈控制功能被转移到云上，而快速反馈回路则保留在本地。然而，由于强调前馈控制，由于控制动作不准确，对机床动态行为建模的不确定性可能会导致故障，从而降低基于云的数控系统的可靠性和性能。该系统将使用测量信号预测故障，并通过在即将发生故障时将控制权从云控制器切换到备用本地控制器，在基于云的冗余数控体系结构中缓解故障。为此，数据驱动的迁移学习框架将使用从连接到基于云的数控系统的其他机器获得的数据来预测并将不确定性降至最低。这种迁移学习利用来自一个来源的数据来学习不同但相关的目标来源。该框架将允许基于云的数控系统：(I)从状态监测信号和过去的故障数据的组合中学习以预测即将到来的故障；(Ii)通过利用状况监测数据来校准其参数是其输入的函数的物理模型来减少不确定性；以及(Iii)规划可行的轨迹，以便在检测到即将到来的故障时从云切换到本地控制器。该项目将解决现有转移学习方法的缺点：(I)结合状态监测和过去的故障数据预测故障事件，以及(Ii)利用状态监测数据的功能参数校准基于物理的模型。这些方法将在CPS试验台上进行实验评估，该试验台由一台3D打印机组成，使用基于云的数控原型从云控制。该奖项反映了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。

Computer numerical control (CNC) is a critical feature of modern manufacturing machines. It provides automated control based on a set of programmed instructions, which traditionally run on a local computer that is physically tethered to the machine. This work envisions a future where manufacturing machines are controlled remotely over the Internet using CNC installed on cloud computers. Among several benefits over traditional CNC, cloud-based CNC holds promise to significantly improve the speed and accuracy of manufacturing machines at low cost. However, a major challenge with cloud-based CNC is that, somewhat like video streaming, it controls manufacturing machines primarily using pre-calculated commands that must be buffered to mitigate Internet transmission delays. For this reason, cloud-based CNC is susceptible to anomalies that result from delayed transmission of information on how the controlled machine is actually behaving. The award supports a scientific investigation into approaches for predicting impending anomalies from data gathered from past experience, and using the predictions to avoid incorrect control actions resulting from inadequate feedback. The U.S. stands to benefit economically from a transition from traditional to cloud-based CNC, since the U.S. is by far the market leader in cloud-based services. The project also will include outreach to U.S. companies, educational curriculum development to increase the U.S. talent pool in manufacturing and data analytics, and activities for middle schoolers in the Detroit area to inspire them to pursue careers in engineering.The objective of the project is to mitigate uncertainties associated with real-time control of manufacturing machines from the cloud using data-driven transfer learning. The knowledge gained will boost the performance of manufacturing machines at low cost by providing the machines with reliable cloud-based CNC. In cloud-based CNC, advanced feedforward control functionalities are transitioned to the cloud while fast feedback loops are retained locally. However, with emphasis on feedforward control, uncertainties in modeling the dynamic behavior of machines could degrade the reliability and performance of cloud-based CNC by causing failures, due to inaccurate control actions. The system will predict failures using measured signals and mitigate them in a redundant, cloud-based CNC architecture by switching control authority from a cloud controller to a back-up local controller in the event of an impending failure. To this end, a data-driven transfer learning framework will predict and minimize uncertainties using data obtained from other machines connected to cloud-based CNC. Such transfer learning leverages data from one source to learn a different, but related, target source. The framework will allow cloud-based CNC to: (i) learn from a combination of condition monitoring signals and past failure data to predict impending failures, (ii) reduce uncertainties by leveraging condition monitoring data to calibrate physical models whose parameters are functions of their inputs, and (iii) plan feasible trajectories for switching from a cloud to a local controller when an impending failure is detected. The project will address the shortcomings of existing transfer learning methods by: (i) predicting failure events from a combination of condition monitoring and past failure data, and (ii) calibration of physics-based models with functional parameters from condition monitoring data. The methods will be evaluated experimentally on a CPS test bed consisting of a 3D printer controlled from the cloud using a cloud-based CNC prototype.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.