PT Unknown AU Krupp, L TI Remaining Machining Tool Life Prediction Using Machine Learning PD 01 PY 2023 DI 10.17185/duepublico/81423 LA en DE Machine Learning; Automatisiertes maschinelles Lernen; Erklärbare künstliche Intelligenz; Freiform-Zerspanung; Individualisierte Produktion; Restlebensdauervorhersage; Werkzeugzustandsüberwachung AB The machinery and equipment manufacturing industry is decisive in achieving a sustainable economy with a savings potential of 37 % of global CO2 emissions. Machining production is a significant factor, accounting for over 15 % of global product development costs. As a result of technological innovation in its application areas, the demands on machining continue to increase, particularly in terms of product quality, flexibility, and component complexity. Examples are the aerospace or tool- and die-making industries, where computer-aided manufacturing of free-form components based on multi-axis machining is standard. At the same time, manufacturing companies are facing the challenges of increasing competition and cost pressure. In order to manufacture at consistently high quality and minimal costs, process and tool monitoring and the subsequent derivation of remaining tool life is of interest. However, due to the increasing customization of production, the prediction of remaining tool life is currently not applicable in the abovementioned areas. Previous process and tool monitoring approaches are too rigid for flexible manufacturing scenarios as they are mainly designed for series production. Accordingly, a methodology for small-batch and single-part production requirements opens up optimization potentials that could not be used so far. Process and tool monitoring methods generally consist of the four components of sensor technology, signal processing and feature extraction, inference of tool and process condition, and prediction of remaining tool life. This work first analyzes the influence of small-batch and single-part production conditions on process and tool monitoring. Mechanical vibration is identified as a particularly suitable monitoring variable. It allows a permanent and process-independent sensor integration without being affected by tool or workpiece adaptations. Based on a physical vibration source model of the machine tool, it is possible to demonstrate the machine independence of the acceleration signals used for vibration sensing. The correlation of the machine-independent signal information with the tool and process state is proven. Thus, the sensor system architecture developed in this work can be used to determine the tool and process state in arbitrary machine tools and flexible machining systems. A limiting factor in previous process and tool monitoring methods is the low adaptivity and explainability of tool and process state determination. Since the methods are designed for series production, the underlying black box models are manually developed once and cannot be adapted or transferred to other applications afterwards. In order to solve this problem, this thesis proposes an explainable and automatically adaptable method for tool and process condition determination. The method uses the automated machine learning approach to enable modeling based on large multivariate sensor datasets without complex feature engineering. An integrated feature evaluation mechanism visually represents the significant features. It is finally possible to determine the tool and process state in a robust, transferable, and performance-optimized way. A 21 % improvement in tool wear prediction error can be achieved. In order to improve the remaining tool life prediction under the random variation of the process conditions in the context of small-batch and single-part production, a new tool and process condition forecasting method was proposed. A reduction of the prediction uncertainty due to the random process variations is achieved by combining temporal machine learning-based models and information about arbitrary future machining operations from process simulations. The method reduces the remaining tool life prediction error by 22 % on average. The remaining tool life prediction is performed with an accuracy of approximately 5 minutes. In relation to the average lifetime of the underlying tools of 85 minutes, the relative error is 6 %. The methods and models developed in this thesis comprehensively extend the applicability of tool and process monitoring to small-batch and single-part production. It is thus possible to perform remaining tool life prediction in flexible manufacturing systems under variable process conditions. The knowledge of the future tool and process conditions enables an optimized production sequence control and cognitive process control to ensure quality and increase productivity. ER