Use of machine learning methods for predicting the Remaining-Useful-Life of tools using the example of mandrel rolls in radial-axial ring rolling

使用机器学习方法以径向-轴向环材轧制中的芯轴辊为例来预测工具的剩余使用寿命

• 批准号: 464881255
• 负责人: Professor Dr.-Ing. Bernd Kuhlenkötter
• 金额: --
• 依托单位: Lehrstuhl für Produktionssysteme
• 依托单位国家: 德国
• 项目类别: Research Grants
• 资助国家: 德国
• 项目状态: 未结题
• 来源: https://gepris.dfg.de/gepris/projekt/464881255?language=en
• 关键词: Use; machine; learning; methods; predicting

For many machines and plants in all branches of industry, seamless ring-shaped components with high requirement specifications, such as high dynamic load capacity and high variability, are required. For the production of such components, radial-axial ring rolling (RARR) is the most important process. With RARR, rings of 100 mm up to 16 m in outer diameter, up to 4 m in ring height and component weights of up to 300 t can be produced. Despite a long research history in ring rolling with many successes, there is still a need for research in the field of predicting mandrel roll fracture. At present, mandrel roll breakage occurs unpredictably and without a directly identifiable cause and can occur up to once per shift, depending on the ring-rolling machine and capacity utilisation. Mandrel roll breakage leads to production downtime, defective rings and unplanned maintenance work, resulting in increased production time and rising costs. As a large number of influencing factors and the non-linear dependencies between them preclude the use of proven investigations to determine qualitative and quantitative influences, as shown in a study by thyssenkrupp Rothe Erde (tkRE) referred to in the preparatory work, machine learning (ML) is used to address this problem. The multidimensional nature of the problem and the non-linear influences make this method particularly suitable, since current algorithms, especially those from the field of deep neural networks, can demonstrate significant success in solving complex problems in a wide range of disciplines. However, the use of ML also creates new challenges, since without a suitable database the use of ML fails. At the same time, data acquisition without domain-specific knowledge is not effective. In this research project, therefore, a plant-independent data acquisition concept for the use of data-driven analysis methods on the one hand, and the foundations for a Remaining-Useful-Life (RUL) model of the mandrel roll in RARR using ML on the other hand are being developed. In order to sufficiently elaborate these two goals, first of all an analysis of the influencing variables for the mandrel roll fracture is carried out. Subsequently, a sensor concept will be developed and extensive sensor technology will be installed at a production plant with a sufficiently large and variable production volume. For a prototypical implementation, the ring roll machine of the Chair of Production Systems can be used within the project, whereby the actual data acquisition is carried out at two industrial companies (tkRE, Schmiedewerke Gröditz). This guarantees a wide range of rolling operations. The data recorded in this way are processed syntactically and semantically and evaluated by means of machine learning methods, so that in the end a regression model for predicting the RUL of a mandrel roll is validated.

对于工业各行各业的许多机器和工厂，要求具有高动载能力和高变化性等高要求的无缝环形件。对于这类零件的生产，径向-轴向环件轧制是最重要的工序。使用RARR，可以生产外径100 mm到16米、环高4米和部件重量300吨的环。尽管环件轧制的研究历史很长，取得了许多成功，但在芯棒轧辊断裂预测方面仍有研究的需要。目前，芯棒轧辊断裂是不可预测的，没有直接可识别的原因，根据轧环机和产能利用率的不同，每个班次最多可能发生一次。芯棒轧辊断裂导致生产停机、钢圈缺陷和计划外维护工作，导致生产时间延长，成本上升。正如蒂森克虏伯公司(ThyssenKrupp Rothe Erde，tkRE)在准备工作中提到的一项研究所表明的那样，由于大量的影响因素及其之间的非线性相关性，无法使用经过验证的调查来确定定性和定量的影响，因此使用机器学习(ML)来解决这一问题。问题的多维性质和非线性影响使该方法特别适合，因为当前的算法，特别是来自深度神经网络领域的算法，可以在解决广泛学科的复杂问题方面表现出巨大的成功。然而，ML的使用也带来了新的挑战，因为如果没有合适的数据库，ML的使用就会失败。同时，没有特定领域知识的数据采集是不有效的。因此，在这个研究项目中，一方面正在开发一种与工厂无关的数据采集概念，以使用数据驱动分析方法，另一方面正在开发使用ML的芯棒轧辊剩余使用寿命(RUL)模型的基础。为了充分阐述这两个目标，首先对芯棒辊断裂的影响因素进行了分析。随后，将开发传感器概念，并将在生产量足够大和可变的生产工厂安装广泛的传感器技术。作为一个原型实施，生产系统主席的环辊机可以在项目中使用，从而在两个工业公司(tkRE、Schmiedewerke Gröditz)进行实际数据采集。这保证了广泛的轧制作业。对记录的数据进行句法和语义处理，并利用机器学习方法进行评估，最终验证了预测芯棒轧辊RUL的回归模型。