A meta-learning approach to select appropriate prognostic methods for the predictive maintenance of digital manufacturing systems
Final Report Abstract
The project resulted in the development of a meta-learning system designed to select appropriate prognostic methods for predictive maintenance based on the state of digitized manufacturing systems. The approach encompasses: (i) a meta-learning model for determining the suitability of prognostic methods for specific manufacturing system scenarios, (ii) an integrated production and maintenance planning method, and (iii) a discrete-event simulation to model the shop floor of a manufacturing system and provide feedback for the evaluation of prognostic methods, the meta-learning model, and the integrated planning method through logistic Key Performance Indicators (KPIs). To expand on these findings, the meta-learning system must be enhanced to incorporate anomaly detection methods, a critical process in developing prognostic methods for predictive maintenance. The literature has addressed the selection of algorithms by meta-learning in dynamic environments; however, there remains a research gap in selecting robust models for anomaly detection in complex scenarios like manufacturing systems. No automated approaches for selecting appropriate methods were found. This leads to the question of how to provide the necessary data for anomaly detection while maintaining the dynamic behavior of the manufacturing system. It is important to explore whether synthetic data generation methods can be utilized for this purpose. Furthermore, there is a need to investigate whether logistic KPIs can be directly utilized as meta-features to effectively enhance the selection of prognostic methods. This extension enables the estimation of the suitability of prognostic methods not only for enhancing machine state forecasting performance but also for understanding the characteristics of the overall manufacturing system. Monitoring KPIs allows for the detection of concept drift in the base learners and the validation tracking of the meta-learner. Regarding the scheduling method, a limitation identified in the project was the computational time and outcomes obtained using the genetic algorithm (GA). A scheduling method that does not rely on dispatching rules could improve the obtained solutions. Recent research has shown that applying machine learning to integrated planning problems can enhance computational performance and results [ZIQI20; SHYA20]. Hence, using Reinforcement Learning (RL) in integrated production and maintenance planning can enhance decision-making effectiveness. The planned adaptations and additions also necessitate corresponding adjustments of the interfaces. This pertains to the service-oriented architecture, which needs to be reviewed and revised. Additionally, the generation of sensor data, the new anomaly detection algorithms, the consideration of human factors in the simulation, and the new planning method introduce new terminologies that should be integrated into the ontology.
Publications
-
Applicability of Algorithm Evaluation Metrics for Predictive Maintenance in Production Systems. 2020 6th IEEE Congress on Information Science and Technology (CiSt), 413-418. IEEE.
Engbers, Hendrik; Alla, Abderrahim Ait; Kreutz, Markus & Freitag, Michael
-
Individual Predictive Maintenance Approach for Diesel Engines in Rail Vehicles. Lecture Notes in Logistics, 236-244. Springer International Publishing.
Engbers, Hendrik; Leohold, Simon & Freitag, Michael
-
Conceptual Model for Integrated Production and Maintenance Planning with Automated Prognostic Method Selection. IFAC-PapersOnLine, 54(1), 635–640.
Engbers, Hendrik; Braghirolli, Lynceo F.; Leohold, Simon; Triska, Yuri; Frazzon, Enzo M. & Freitag, Michael
-
Prognostic Methods for Predictive Maintenance: A generalized Topology. IFAC-PapersOnLine, 54(1), 629-634.
Leohold, Simon; Engbers, Hendrik & Freitag, Michael
