Simulation-based deep reinforcement learning for modular production systems. - In: 2020 Winter Simulation Conference (WSC), (2020), S. 1596-1607
Modular production systems aim to supersede the traditional line production in the automobile industry. The idea here is that highly customized products can move dynamically and autonomously through a system of flexible workstations without fixed production cycles. This approach has challenging demands regarding planning and organization of such systems. Since each product can define its way through the system freely and individually, implementing rules and heuristics that leverage the flexibility in the system in order to increase performance can be difficult in this dynamic environment. Transport tasks are usually carried out by automated guided vehicles (AGVs). Therefore, integration of AI-based control logics offer a promising alternative to manually implemented decision rules for operating the AGVs. This paper presents an approach for using reinforcement learning (RL) in combination with simulation in order to control AGVs in modular production systems. We present a case study and compare our approach to heuristic rules.
https://doi.org/10.1109/WSC48552.2020.9384089
A method for predicting high-resolution time series using sequence-to-sequence models. - In: 2020 Winter Simulation Conference (WSC), (2020), S. 1075-1086
With the increasing availability of data, the desire to interpret that data and use it for behavioral predictions arises. Traditionally, simulation has used data about the real system for input data analysis or within data-driven model generation. Automatically extracting behavioral descriptions from the data and representing it in a simulation model is a challenge of these approaches. Machine learning on the other hand has proven successful to extract knowledge from large data sets and transform it into more useful representations. Combining simulation approaches with methods from machine learning seems therefore promising to combine the strengths of both approaches. Representing some aspects of a real system by a traditional simulation model and others by a model incorporating machine learning, a hybrid system model (HSM) is generated. This paper suggests a specific HSM incorporating a deep learning method for predicting the anticipated power usage of machining jobs.
https://doi.org/10.1109/WSC48552.2020.9383969
On the usage of deep learning for modelling energy consumption in simulation models. - In: Simulation Notes Europe, ISSN 2164-5353, Bd. 30 (2020), 4, S. 165-174
With the increasing availability of data, the desire to interpret that data and use it for behavioral predictions arises. Traditionally, simulation has used data about the real system for input data analysis or within data-driven model generation. Automatically extracting behavioral descriptions from the data and representing it in a simulation model is a challenge for these approaches. Machine learning on the other hand has proven successful in extracting knowledge from large data sets and transforming it into more useful representations. Combining simulation approaches with methods from machine learning seems, therefore, promising. Representing some aspects of a real system by a traditional simulation model and others by a model generated from machine learning, a hybrid system model (HSM) is generated. This paper discusses such HSMs and suggests a specific HSM incorporating a deep learning method for predicting the power consumption of machining jobs.
https://doi.org/10.11128/sne.30.tn.10536
Knowledge discovery in simulation data. - In: ACM transactions on modeling and computer simulation, ISSN 1558-1195, Bd. 30 (2020), 4, S. 24:1-24:25
This article provides a comprehensive and in-depth overview of our work on knowledge discovery in simulations. Application-wise, we focus on manufacturing simulations. Specifically, we propose and discuss a methodology for designing, executing, and analyzing large-scale simulation experiments with a broad coverage of possible system behavior targeted at generating knowledge about the system. Based on the concept of data farming, we suggest a two-phase process which starts with a data generation phase, in which a smart experiment design is used to set up and efficiently execute a large number of simulation experiments. In the second phase, the knowledge discovery phase, data mining and visually aided analysis methods are applied on the gathered simulation input and output data. This article gives insights into this knowledge discovery phase by discussing different machine learning approaches and their suitability for different manufacturing simulation problems. With this, we provide guidelines on how to conduct knowledge discovery studies within the manufacturing simulation context. We also introduce different case studies, both academic and applied, and use them to validate our methodology.
https://doi.org/10.1145/3391299
Sequence to Sequence Modelle zur hochaufgelösten Prädiktion von Stromverbrauch. - In: Proceedings ASIM SST 2020, (2020), S. 149-157
Modelling power consumption for jobs on a ma-chine tool is commonly performed by measuring the real power consumption of comparable jobs and machines. The so gathered data is then processed to represent the time-av-eraged sums of power consumptions of previous jobs. These values of power consumption are then used for upcoming comparable jobs. This approach allows for no high-resolution prediction of power consumption and further presumes static processing times of jobs. Here we propose a new approach to model power consumption that incorporates a Sequence-to-Sequence model, which generates time series according to dynamic data, that describes a numerical control code and environment settings such as state of tools, etc.
https://doi.org/10.11128/arep.59.a59021
Concept for the comparison of intralogistics designs with real factory layout using augmented reality, SLAM and marker-based tracking. - In: 53rd CIRP Conference on Manufacturing Systems 2020, (2020), S. 341-346
In the automotive industry, the intralogistics planning faces the problem of matching the planning data with the current conditions in the assembly hall. The large variety of parts leads to a constantly changing production. Based on this, we establish an approach for the comparison using augmented reality (AR) and simultaneous localization and mapping (SLAM). The use of SLAM enables the consistent application of AR in an assembly hall. Based on this, the objective of this article is to visualize 3D objects from the corresponding CAD planning tool in the real factory and thus the comparison of the intralogistics design with the real factory is possible due to AR. Nevertheless, there was a lack of practical implementations in intralogistics and therefore the concept is evaluated by two prototypical solutions. The first one is implemented on an iPhone 7 using SLAM. The second prototype is developed on a HoloLens 2 and is based on a hybrid tracking solution, SLAM and marker tracking.
https://doi.org/10.1016/j.procir.2020.03.039
A method for robustness optimization using generative adversarial networks. - In: SIGSIM-PADS '20, (2020), S. 1-10
This paper presents an approach for optimizing the robustness of production and logistic systems based on deep generative models, a special method of deep learning. Robustness here refers to setting controllable factors of a system in such a way that variance in the uncontrollable factors (noise) has a minimal effect on given output parameters. In a case study, the proposed method is tested and compared to a traditional method for robustness analysis. The basic idea is to use deep neural networks to generate data for experiment plans and rate them by use of a simulation model of the production system. We propose to use two Generative Adversarial Networks (GANs) to generate optimized experiment plans for the decision factors and the noise factors, respectively, in a competitive, turn-based game. In one turn, the controllable factors are optimized and the noise remains constant, and vice versa in the next turn. For the calculations of the robustness, the planned experiments are conducted and rated using a simulation model in each learning step.
https://doi.org/10.1145/3384441.3395981
Automatische Modellgenerierung - Stand, Klassifizierung und ein Anwendungsbeispiel. - In: Ablaufsimulation in der Automobilindustrie, (2020), S. 333-347
Die automatische Modellgenerierung (AMG) ist ein Ansatz, der darauf abzielt, sowohl die Aufwände einer Simulationsstudie zu senken als auch die Qualität der erzeugten Modelle zu verbessern. Unter automatischer Modellgenerierung werden im Kontext der Simulation verschiedene Ansätze subsumiert, die es erlauben, Simulationsmodelle oder zumindest Teile von Simulationsmodellen mittels Algorithmen zu erzeugen. Eine umfassende Klassifizierung der Ansätze nach verschiedenen Merkmalen ist Ausgangspunkt weiterer Betrachtungen des Beitrags, in denen u. a. verschiedene technische Ansätze zur Modellgenerierung diskutiert werden. Weiterhin werden ergänzende Techniken, die die eigentliche Modellgenerierung flankierenden, wie z. B. die automatische Modellinitialisierung, diskutiert. Als ein möglicher Lösungsansatz wird beispielhaft ein Framework zur automatischen Modellgenerierung, -initialisierung und -adaption, welches das standardisierte Core Manufacturing Simulation Data (CMSD) Format als Basis nutzt, beschrieben.
https://doi.org/10.1007/978-3-662-59388-2_23
Wissensentdeckung im Kontext der Produktionssimulation. - Ilmenau : Universitätsverlag Ilmenau, 2020. - 1 Online-Ressource (XII, 217, XIV-XX Seiten)
Technische Universität Ilmenau, Dissertation 2019
Die diskrete Simulation stellt eine wichtige und etablierte Methode zur Untersuchung des dynamischen Verhaltens von komplexen Produktions- und Logistiksystemen dar. Sie ist daher zur Planung, Steuerung und Kontrolle solcher Systeme unerlässlich, beispielsweise in der Automobilindustrie oder in der Halbleiterfertigung. Klassische Simulationsstudien zielen in diesem Kontext üblicherweise darauf ab, typische, vorab definierte Fragestellungen zu beantworten. Dies geht oftmals einher mit der Simulation und Analyse einiger weniger vorab definierter Szenarien. Wirkzusammenhänge, die über diesen definierten Projektrahmen hinausgehen, bleiben daher eventuell unentdeckt. Auf der anderen Seite erwachsen mit steigender Rechenleistung und der allgemeinen Verfügbarkeit von Big-Data-Infrastrukturen neue Möglichkeiten zur Durchführung von sehr großen Bandbreiten von Simulationsexperimenten, um das Verhalten des Modells möglichst vollständig abzudecken und automatisiert auszuwerten. Dies wird allgemein als Data Farming bezeichnet. Ziel dieser Arbeit war es, die Methode des Data Farming für die Nutzung zur Wissensentdeckung in Produktionssimulationen zu übertragen und weiterzuentwickeln. Dazu wurde ein ganzheitliches Konzept ausgearbeitet, um unbekannte, versteckte und potenziell nützliche Wirkzusammenhänge in großen Mengen von Simulationsdaten entdecken zu können. Das Konzept beinhaltet hierzu die Auswahl geeigneter Experimentdesignmethoden, die Anwendung und Ausgestaltung von geeigneten Data-Mining-Verfahren in einem dafür zweckmäßigen und zielgerichteten Analyseprozess sowie die Definition geeigneter Visualisierungs- und Interaktionsmethoden zur iterativen, anwenderorientierten Analyse großer Mengen von Simulationsdaten. Darüber hinaus wurde das Konzept in einem ganzheitlichen Softwareframework prototypisch implementiert. Die Anwendbarkeit des Konzeptes wurde anhand von vier Fallstudien aufgezeigt und validiert. Die Fallstudien beinhalteten hierbei zwei akademische Laborstudien sowie zwei Industrieanwendungsfälle.
https://www.db-thueringen.de/receive/dbt_mods_00040526
Visualization and interaction for knowledge discovery in simulation data. - In: Hawaii International Conference on System Sciences 2020, (2020), S. 1340-1349
Discrete-event simulation is an established and popular technology for investigating the dynamic behavior of complex manufacturing and logistics systems. Besides traditional simulation studies that focus on single model aspects, data farming describes an approach for using the simulation model as a data generator for broad scale experimentation with a broader coverage of the system behavior. On top of that we developed a process called knowledge discovery in simulation data that enhances the data farming concept by using data mining methods for the data analysis. In order to uncover patterns and causal relationships in the model, a visually guided analysis then enables an exploratory data analysis. While our previous work mainly focused on the application of suitable data mining methods, we address suitable visualization and interaction methods in this paper. We present those in a conceptual framework followed by an exemplary demonstration in an academic case study.
https://doi.org/10.24251/HICSS.2020.165