From simulation to injection molding machine - Optimized process setup based on machine learning Prof. Dr.-Ing. Christian Hopmann; Julian Heinisch, M.Sc. 2017 International Mold Conference Bucheon 3 rd November 2017 From simulation to injection molding machine – Optimized process setup based on machine learning 1 Dr. Feistkorn AG Executive Consulting Dr. Begemann Das Bild kann nicht angezeigt werden. mrink-consulting OPC – Orth Plastics Consulting Institute of Plastics Processing (IKV) Institute and Chair Founded in 1950 by industry and the skilled crafts, since then supported by an association of sponsors Integrated into the network and infrastructure of RWTH Aachen University as an affiliated institute Industrial network More than 290 members across all industries (thereof 1/3 from abroad) Material producers Plastics processors, OEMs Machine manufacturers Service providers Departments Injection Molding Extrusion and Rubber Technology Composites and Polyurethanes Part Design and Materials Technology Centre for Analysis and Testing of Plastics Training and Further Education Employees at IKV Institute director: Prof. Hopmann 80 scientists 50 technical/administrative staff More than 200 student researchers From simulation to injection molding machine – Optimized process setup based on machine learning 2 February 28 – March 1, 2018 International Colloquium Plastics Technology From simulation to injection molding machine – Optimized process setup based on machine learning 3 Outline Motivation injection molding setup in a digitized industry Machine learning for process setup in injection molding Injection molding experiments and simulations as database Modeling results Combined models using simulation and experimental data Optimization based of process settings Conclusion and outlook From simulation to injection molding machine – Optimized process setup based on machine learning 4 The production becomes more: flexible individualized Self-optimizing New Challenges: How do future business models and strategies look like? How can different technical components be integrated? How can different organizational levels be integrated? What is the role of humans in a future production? Motivation: Aspects of Industry 4.0 Linking technological production systems to digital business models Linking the data layer with physical processes Exchange and processing of relevant information in real-time Integrating all units of the value chain (machines, molds, raw materials, products, employees, design and development) by use of virtual shadows PLM = Product Lifecycle Management Security & safety Methods of Industry 4.0 Integrated engineering PLM Decentral production control Big data Human machine interaction Horizontal integration Vertical integration Real time From simulation to injection molding machine – Optimized process setup based on machine learning 5 Process setup in injection molding The conventional trial and error approach Objective Finding setting parameters which deliver the desired quality output while ensuring the best possible process robustness and productivit y with the lowest effort possible Challenges Integration of manual and experience-driven processes in a virtual shadow Trained staff is required Subjective result, no systemization Setup is aborted after the minimum quality target is reached without further optimization Machine settings Product quality and process productivity Machine Operator Trial of setting parameters based on: Recommendations of material supplier Experience, knowhow (and intuition) of the operator Quality feedback From simulation to injection molding machine – Optimized process setup based on machine learning 6 Outline Motivation injection molding setup in a digitized industry Machine learning for process setup in injection molding Injection molding experiments and simulations as database Modeling results Combined models using simulation and experimental data Optimization based of process settings Conclusion and outlook From simulation to injection molding machine – Optimized process setup based on machine learning 7 Process setup by means of machine learning – Approach based on experimental data IM-Experiments Design of experiments Modeling Optimization Optimized process settings From simulation to injection molding machine – Optimized process setup based on machine learning 8 Simulations IM-Experiments Process setup by means of machine learning – Combined learning approach within the Cluster of Excellence IM-Experiments Design of experiments Modeling Optimization Optimized process settings Correction of the inevitable gap between simulation and experiments Reduced number of pratical trials From simulation to injection molding machine – Optimized process setup based on machine learning 9 Logistic regressions Support Vector Machines Methods of machine learning Connectivity- based clustering Distribution- based clustering Classification Regression Clustering Supervised learning Unsupervised learning Machine learning Linear regression Support Vector Machine Gaussian Process Regression Decision trees Artificial neural networks From simulation to injection molding machine – Optimized process setup based on machine learning 10 ∑ ... 𝑥𝑥 2 𝑥𝑥 1 ... 𝑥𝑥 i 𝑤𝑤 1j 𝑤𝑤 2j 𝑤𝑤 ij 𝝋𝝋 𝑦𝑦 1 Weights Activation threshold θ j Output Transfer function Inputs Principle and application of artificial neural networks (ANN) Output layer Hidden layer Input layer 𝑤𝑤 ij 𝑤𝑤 jk 𝑥𝑥 1 𝑥𝑥 i 𝑥𝑥 2 ... 𝑦𝑦 1 𝑦𝑦 k 𝑦𝑦 2 ... Typical applications of ANN Data mining Image processing Speech recognition Weather forecasts Modeling of technical systems Control engineering From simulation to injection molding machine – Optimized process setup based on machine learning 11 Outline Motivation injection molding setup in a digitized industry Machine learning for process setup in injection molding Injection molding experiments and simulations as database Modeling results Combined models using simulation and experimental data Optimization based of process settings Conclusion and outlook From simulation to injection molding machine – Optimized process setup based on machine learning 12 Quality values: Plate specimens Box specimens For both specimens: Dimension a and Dimension b Part weight Input and output variables from simulation and practical injection molding trials b a Varied parameters: Mold temperature Melt temperature Packing time Packing pressure Cooling time Injection speed Central composite design (CCD) of experiments: Full-factorial experimental design (64-point cube) Star points outside the cube (12 points) Central point a From simulation to injection molding machine – Optimized process setup based on machine learning 13 Exemplary comparison of simulation and real data – Box part weights Mittelwert, SG-Versuche Simulation Box specimen: Simulation and experimental data 50,0 51,0 52,0 53,0 54,0 55,0 56,0 57,0 0 10 20 30 40 50 60 70 Box weight [g] Test point [-] Box, measured Box, simulated From simulation to injection molding machine – Optimized process setup based on machine learning 14 Exemplary comparison of simulation and real data – Plate dimensions Plate specimen: Simulation and experimental data 87,9 88,4 88,9 89,4 89,9 135 136 137 138 139 140 0 20 40 60 Plate width [mm] Plate length [mm] Test point [-] Plate length, measured Plate length, simulated From simulation to injection molding machine – Optimized process setup based on machine learning 15 Effect analysis based on simulation data enables a reduction of the experimental design -100 -50 0 50 100 150 200 Standardised effect [-] Parameter [-] Plate weight Plate weight simulated Box weight Box weight simulated Effect analysis based on simulation data shows only smalls effects of cooling time and injection speed Therefore, the realization of targeted weights or dimensions could be guaranteed without these 2 parameters A variation of the 2 parameters could consequently be omitted in the experimental trials For a central composite design the necessary test points could be reduced from 77 (for 6 parameters) to 25 (for 4 parameters From simulation to injection molding machine – Optimized process setup based on machine learning 16 Challenges for learning methods and assessment of prediction quality Preferably as little data as possible for model generation Unavoidable differences between experimental and simulation data Higher information content in real data Assessment of prediction quality: Mean squared error (MSE): 𝑀𝑀𝑀𝑀𝑀𝑀 = 1 𝑛𝑛 � 𝑖𝑖=1 𝑛𝑛 ( � 𝑦𝑦 𝑖𝑖 − 𝑦𝑦 𝑖𝑖 ) ² 𝑛𝑛 : Number of predicted quality values � 𝑦𝑦 𝑖𝑖 : Predicted quality value i 𝑦𝑦 𝑖𝑖 : Actual quality value i Unless otherwise stated, calculated based on 20 % of test data Quality value Cycle [-] μ Likelihood density [-] From simulation to injection molding machine – Optimized process setup based on machine learning 17 Outline Motivation injection molding setup in a digitized industry Machine learning for process setup in injection molding Injection molding experiments and simulations as database Modeling results Combined models using simulation and experimental data Optimization based of process settings Conclusion and outlook From simulation to injection molding machine – Optimized process setup based on machine learning 18 0,01 0,1 1 10 0 50 100 150 200 MSE [-] Training cycle [-] Trainingsperformance Testperformance Exemplary training progress using Levenberg Marquardt Backpropagation (LMBP) algorithm Box specimen Simulation data (weight and dimensions) 80 % Training, 20 % Test data ANN: 6 Neurons, LMBP Trainings performance Test performance From simulation to injection molding machine – Optimized process setup based on machine learning 19 Variations due to initialization values of the neuron weights and computing times The random initial values of the neuron weights cause variations for the prediction accuracy of ANNs trained with the same data 100 ANN a trained simulatanosly and are averaged Training of 100 ANN takes 15 s on a conventional personal computer 0 0,1 0,2 0,3 0,4 0,5 0 50 100 MSE [-] Network output [-] Box and plate specimen Experimental data 80 % Training, 20 % Test data ANN: 3-6 Neurons, LMBP