Machine Learning for Cyber Physical Systems: Selected papers from the International Conference ML4CPS 2016 PDF
Preview Machine Learning for Cyber Physical Systems: Selected papers from the International Conference ML4CPS 2016
Technologien für die intelligente Automation Technologies for Intelligent Automation Jürgen Beyerer Oliver Niggemann Christian Kühnert Editors Machine Learning for Cyber Physical Systems Selected papers from the International Conference ML4CPS 2016 Technologien für die intelligente Automation Technologies for Intelligent Automation Band 3 Weitere Bände in dieser Reihe http://www.springer.com/series/13886 Ziel der Buchreihe ist die Publikation neuer Ansätze in der Automation auf wissenschaft lichem Niveau, Themen, die heute und in Zukunft entscheidend sind, für die deutsche und internationale Industrie und Forschung. Initiativen wie Industrie 4.0, Industrial Internet oder Cyberphysical Systems machen dies deutlich. Die Anwendbarkeit und der indust rielle Nutzen als durchgehendes Leitmotiv der Veröffentlichungen stehen dabei im Vorder grund. Durch diese Verankerung in der Praxis wird sowohl die Verständlichkeit als auch die Relevanz der Beiträge für die Industrie und für die angewandte Forschung gesichert. Diese Buchreihe möchte Lesern eine Orientierung für die neuen Technologien und deren Anwendungen geben und so zur erfolgreichen Umsetzung der Initiativen beitragen. Herausgegeben von inIT – Institut für industrielle Informationstechnik Hochschule OstwestfalenLippe Lemgo, Germany Jürgen Beyerer · Oliver Niggemann Christian Kühnert (Eds.) Machine Learning for Cyber Physical Systems Selected papers from the International Conference ML4CPS 2016 Editors Jürgen Beyerer Christian Kühnert Karlsruhe, Germany Karlsruhe, Germany Oliver Niggemann Lemgo, Germany Technologien für die intelligente Automation ISBN 9783662538050 ISBN 9783662538067 (eBook) DOI 10.1007/9783662538067 Library of Congress Control Number: 2016955525 Springer Vieweg © SpringerVerlag GmbH Germany 2017 This work is subject to copyright. All rights are reserved by the Publisher, whether the whole or part of the material is concerned, specifically the rights of translation, reprinting, reuse of illustrations, recitation, broadcasting, reproduction on microfilms or in any other physical way, and transmission or information storage and retrieval, electronic adaptation, computer software, or by similar or dissimilar methodology now known or hereafter developed. The use of general descriptive names, registered names, trademarks, service marks, etc. in this publication does not imply, even in the absence of a specific statement, that such names are exempt from the relevant protective laws and regulations and therefore free for general use. The publisher, the authors and the editors are safe to assume that the advice and information in this book are believed to be true and accurate at the date of publication. Neither the publisher nor the authors or the editors give a warranty, express or implied, with respect to the material contained herein or for any errors or omissions that may have been made. Printed on acidfree paper This Springer Vieweg imprint is published by Springer Nature The registered company is SpringerVerlag GmbH Germany The registered company address is: Heidelberger Platz 3, 14197 Berlin, Germany Preface Cyber Physical Systems are characterized by their ability to adapt and to learn. Theyanalyzetheirenvironment,learnpatterns,andtheyareabletogeneratepre- dictions. Typical applications are condition monitoring, predictive maintenance, image processing and diagnosis. Machine Learning is the key technology for these developments. ThesecondconferenceonMachineLearningforCyber-Physical-SystemsandIndus- try 4.0 - ML4CPS - was held at the Fraunhofer IOSB in Karlsruhe, on September 29th 2016. The aim of the conference is to provide a forum to present new ap- proaches, discuss experiences and to develop visions in the area of data analysis forcyber-physicalsystems.Thisbookprovidestheproceedingsofallcontributions presented at the ML4CPS 2016. The editors would like to thank all contributors that led to a pleasant and re- warding conference. Additionally, the editors would like to thank all reviewers for sharing their time and expertise with the authors. It is hoped that these proceed- ings will form a valuable addition to the scientific and developmental knowledge in the research fields of machine learning, information fusion, system technologies and industry 4.0. Prof. Dr.-Ing. Jürgen Beyerer Dr.-Ing. Christian Kühnert Prof. Dr.-Ing. Oliver Niggemann Table of Contents Page A Concept for the Application of Reinforcement Learning in the Optimization of CAM-Generated Tool Paths ........................... 1 Caren Dripke, Sara Höhr, Akos Csiszar, Alexander Verl Semantic Stream Processing in Dynamic Environments Using Dynamic Stream Selection.................................................... 9 Michael Jacoby and Till Riedel Dynamic Bayesian Network-Based Anomaly Detection for In-Process Visual Inspection of Laser Surface Heat Treatment ...................... 17 Alberto Ogbechie, Javier Díaz-Rozo, Pedro Larrañaga, Concha Bielza A Modular Architecture for Smart Data Analysis using AutomationML, OPC-UA and Data-driven Algorithms ................................. 25 Christian Kühnert, Miriam Schleipen, Michael Okon, Robert Henßen, Tino Bischoff Cloud-based event detection platform for water distribution networks using machine-learning algorithms..................................... 35 Thomas Bernard, Marc Baruthio, Claude Steinmetz, Jean-Marc Weber A Generic Data Fusion and Analysis Platform for Cyber-Physical Systems . 45 Christian Kühnert, Idel Montalvo Arango Agent Swarm Optimization: Exploding the search space ................. 55 Idel Montalvo Arango, Joaquín Izquierdo Sebastián Anomaly Detection in Industrial Networks using Machine Learning........ 65 Ankush Meshram, Christian Haas A Concept for the Application of Reinforcement Learning in the Optimization of CAM-Generated Tool Paths Dripke, Caren1, Höhr, Sara1, Csiszar, Akos2, and Verl, Alexander1 1 Institute for Control Engineering of Machine Tools and Manufacturing Units (ISW) Seidenstr. 36, Stuttgart, 70174, Germany, Tel. +49-711-685 84500 [email protected] 2 Graduate School of Excellence advanced Manufacturing Engineering, University of Stuttgart, 70569 Stuttgart, Germany Abstract. Cyber physical systems (CPS) are changing the way machine tools function and operate. As the CAD-CAM-CNC tool chain gains in- telligence the boundaries of the elements of the tool chain become blurred and new features, based on advancements in artificial intelligence can be integrated. The main task of the CAD-CAM-CNC chain is to generate the cutter trajectories for the manufacturing operation. Driven by sustainabil- ityandtheneedforcapacity,theneedarisestooptimizethepathsthrough this tool chain. In this paper a concept for path optimization with rein- forcement learning is proposed, with focus on the reward function, specific to tool path optimization via the channel method. Keywords: CAD-CAM-CNC-Chain, Reinforcement Learning, Tool path, Smoothing, Channel Method 1 Introduction Machine learning in the field of manufacturing engineering is not only applicable to classical planning problems such as production planning or logistics, but can be used to optimize detailed aspects of the manufacturing process as well. The manufacturingprocessitself,aswellastheengineeringphaseofthemanufacturing process of milled parts, is a tedious, time-consuming activity. The main task is the definition of cutting tool paths. This is done in the CAM (Computer Aided Manufacturing) part of the well-known CAD-CAM-CNC chain. The control of the machinetoolbytheComputerizedNumericalControl(CNC)is,asofnow,defined by the CAD-CAM part and has the task of moving the machine axes in such a way, that the cutting tool follows the path generated by the CAM tool. Clear boundaries between different components of the CAD-CAM-CNC chain become blurred, as each component gains intelligence. Machine tools are more and more often viewed as Cyber Physical Production Systems (cPPS) [1]. It is desired, that a machine tool, acting as a CPS, not just executes the path (program) which is given to it, but optimizes this, based on its capabilities, prior to execution. As an overall concept we propose a reinforcement learning approach, based on which the machine tool learns to optimize the paths it receives. As this is an ongoing © Springer-Verlag GmbH Germany 2017 J. Beyerer et al. (Eds.), Machine Learning for Cyber Physical Systems, Technologien für die intelligente Automation 3, DOI 10.1007/978-3-662-53806-7_1 2 Dripke et al. Fig.1: Surface imperfections with non-optimized jittered tool path. research in its preliminary stage, this paper is focused on the reward function of the reinforcement learning approach. Thepaperisstructuredasfollows:Thenextchapteroffersinsightstotheproblem. Afterwards the related works are presented. Chapter 4 gives a brief overview of the theoretical background and Chapter 5 presents our concept and discussion of its preliminary evaluations. Finally, conclusions are drawn and a plan for future actions is presented. 2 Effects of Non-Optimal CAM-generated Tool Paths The CAM-generated paths are often not ideal. The reason behind this lies in the differences between modeling of NURBS (Non-Uniform Rational B-Spline) and other polynomial-based surfaces in the CAD tool and discretizing these surfaces by using a simple grid in the CAM tool. The grid is neither correlated with the CAD models manufacturing tolerances nor with the splines used in the CNC to createcontinuousmotionprofiles.Correlatingtheseaspectsbyhandtakestimeand requirestheexpertiseandexperienceofaskilledCAMengineer.Ifleftuncorrelated the cutting velocity will not be as high as it could be. These tool paths lead to changesinthefeedrateandthuscauseadegradedsurfacequality(seeFigure1)and lengthier then required machining times. Currently, it is only feasible for large lot sizestoundertakesuchoptimizationworkmanually.AnintelligentCNCshouldbe abletodetectsuboptimalpartsoftoolpathsindependentlyoflotsizeandoptimize these. Using reinforcement learning to carry out this optimization problem seems promising as it has the potential to conserve the knowledge generated (in contrast tooptimizationmethods).Inthispaper,thefirststepsoptimizingCNCtrajectories via reinforcement learning are presented. We present a step-wise validation of the different proposed elements in this phase of the ongoing research. Currently, the reward function is in our focus. A given tool path segment is evaluated by the proposed reward function. The path is then altered by moving the control points, as it will be later done by the reinforcement algorithm, and it is then reevaluated to see if the reward function reflects the changes in path quality. 3 Related Work In the scientific literature many examples can be found where machine learning has been used for improving manufacturing or trajectory generation related activ- ities, but not specifically for the improvement of already existent trajectories for manufacturing purposes. In [2] supervised learning with artificial neural networks A Concept for the Application of Reinforcement Learning 3 areproposedtopredictthestabilityofthecuttingprocess.In[3]supervisedlearn- ing and artificial neural networks are used to predict the surface quality of the work piece for a given cutting process (milling and turning). These do not include improvements for the manufacturing processes, although predicting the quality is always a first step towards improvement. Related to paths or trajectories, machine learning applications can be found in numerous papers dealing with trajectory following for robot arm. In most cases, like in [4,5], machine learning, mostly su- pervised learning, is proposed to make a robot arm learn to follow a predefined trajectory as accurately as possible. The focus lies in eliminating the need for the inverse kinematics transformation as opposed to the smoothness of the obtained contours. Reinforcement learning has also been used related to robotic path plan- ning. In [6] a concept is proposed where a robot learns its task instead of being pre-programmed. The paper is focused on corridor-following and reaching desti- nation points, and not so much on the trajectory itself. Optimization or similar methods have been previously used to optimize trajectories for machining. Exam- plesforthesecanbefoundin[7,8,9,10].Theseworksoptimizethetrajectoriesusing the same general characteristic, which we are also aiming for. However, optimiza- tion methods have the inherent drawback that these will not learn over time. Our distant goal is to exploit the characteristic of reinforcement learning, which allows a faster run time after a learning stage. 4 Tool Path Optimization with Reinforcement Learning via the Channel Method In this chapter we present a brief overview of the underlying concepts of our ap- proach to tool path optimization using reinforcement learning. By knowing the manufacturing tolerance of the work piece (always known prior to manufacturing), a channel (similar to a tunnel) is formed around the contour that is cut by the tool (see Figure 2) [11]. As long as the tool path stays in this tunnel, way points can be moved around, inserted and sometimes even deleted freely. This property is exploited in the proposed concept. By moving, inserting and deleting way points, a post processor based on reinforcement learning techniques can make sure that the changes in the curvature of the tool path are optimal and bounded by respecting the tolerances. Bézier curves are an approximation method for the generation of smooth curves andsurfaces,thathasbeenpresentedbyBézierinthe1960s[12].Hisobjectivewas to create an intuitive concept to change a curve’s appearance by manipulation of the defining control points. The Bézier approximation curve is defined by (cid:2)n P(t)= PiBn,i, 0<t<1 (1) i=0 where Bn,k are the Bernstein Polynomials (cid:3) (cid:4) n Bn,k = k tk(1−t)n−k (2) There are multiple properties of the Bézier curve that make them well suited for our application [13, Ch. 6.4]:
The list of books you might like
