A meta-learning approach to select appropriate prognostic methods for the predictive maintenance of digital manufacturing systems

为了保证机器的持续可用性和避免高昂的维修成本，维护计划对制造公司来说是至关重要的。过去几年的技术发展，可以概括为工业4.0和先进数字化，为实时描述数字化制造系统的状态提供了新的潜力。这为从传统的被动或定期维护计划转向更有效的基于状态和预测性维护提供了基础。然而，预测机器和部件未来失效行为的方法的适用性和性能在很大程度上取决于机器、部件及其配置的状态。因此，适用性可以随着时间的推移而改变。尽管在过去的几年中对预测性维护进行了大量的研究，但在根据制造系统及其机器的当前状态选择适当的预测性维护预测方法的方法方面存在明显的研究差距。这个联合研究项目的目标是开发一个元学习系统，根据机器的传感器数据，根据制造系统的当前状态，选择合适的预测方法进行预测性维护。该项目将由三个研究小组联合进行，分别是：(1)德国不来梅大学BIBA - Bremer农业生产与物流研究所，(2)巴西弗洛里亚诺波利斯圣卡塔琳娜联邦大学，(3)巴西阿雷格里港南巴西格兰德州联邦大学，这三个研究小组的研究资料是互补的。BIBA的研究小组将开发一种元学习方法，根据历史和传感器数据选择合适的预测方法。该研究组将开发一种综合生产和维修计划方法。UFRGS的研究小组将开发一个面向服务的体系结构，以确保在制造系统和预测以及规划方法之间进行适当的数据交换。整合这些不同的模块，提议项目的结果将是一个元学习预测维护系统，该系统将能够(i)使用数字化制造系统的传感器数据实时描述系统状态，（ii）根据机器及其组件的状态动态选择合适的预测方法。（iii）计算综合生产和维护计划；（iv）通过预测误差评估所选预测方法的性能以及随后的计划决策，此外，通过使用物流关键绩效指标，如机器利用率和吞吐量时间。通过这种方式，元学习预测性维护系统将帮助制造企业实现更好的生产和维护计划，并提高生产绩效。

Maintenance planning is of paramount importance for manufacturing companies in order to assure a constant availability of their machines and to avoid high repair costs. The technological developments of the last years, which can be summarized under the terms Industry 4.0 and Advanced Digitalization, offer new potentials to describe the state of a digitalized manufacturing system in real-time. This provides a basis for moving from classical reactive or periodic maintenance planning to more efficient condition-based and predictive maintenance. However, the suitability and performance of prognostic methods to predict the future failure behavior of machines and components depends strongly on the state of a machine, its components and their configuration. Hence, the suitability can change over time. Despite the large amount of research regarding predictive maintenance in the last years, there is a clear research gap regarding a method to select appropriate prognostic methods for predictive maintenance depending on the current state of a manufacturing system and its machines.The objective of this joint research project is to develop a meta-learning system that selects suitable prognostic methods for predictive maintenance depending on the current state of a manufacturing system using sensor data of the machines. The project will be jointly conducted by three research groups at (i) BIBA - Bremer Institut für Produktion und Logistik at the University of Bremen, Germany, (ii) the Federal University of Santa Catarina (UFSC), Florianopolis, Brazil, and (iii) the Federal University of Rio Grande do Sul (UFRGS), Porto Alegre, Brazil, who have complementary research profiles. The research group at BIBA will develop a meta-learning method to select suitable prognostic methods based on historical and sensor data. The research group at UFSC will develop an integrated production and maintenance planning method. The research group at UFRGS will develop a service-oriented architecture to assure an appropriate data-exchange between a manufacturing system and the prognostic as well as the planning method. Integrating these different modules, the result of the proposed project will be a meta-learning predictive maintenance system that will be capable to (i) use sensor data of digitalized manufacturing systems to describe the system state in real-time, (ii) select suitable prognostic methods dynamically based on the state of a machine and it’s components, (iii) compute an integrated production and maintenance plan and (iv) evaluate the performance of the selected prognostic methods as well as the subsequent planning decisions by their forecast errors and, in addition, by using logistic key performance indicators such as machine utilization and throughput times. In this way, the meta-learning predictive maintenance system will help a manufacturing company to achieve a better production and maintenance planning and an increased production performance.