Industrial Transfer Learning
Introduction to Industrial Transfer Learning
Industrial Transfer Learning Motivation Machine Learning in Manufacturing Decision Support Automation Process Control Self-Optimization Predictive Quality Predictive Maintenance Challenges for machine learning in manufacturing ! Dynamic processes → high training effort Insufficient data → representative and reliable data required Industrial Transfer Learning 3 Chair of Technologies and Management of Digital Transformation, University of Wuppertal
Industrial Transfer Learning Challenges for Machine Learning in Production ▪ One key requirement of successful ML: representative and reliable data basis ▪ Main data sources in production have advantages and disadvantages regarding costs and data quantity Running Production Test Environment Simulation (Experiments) High quantity Small quantity Simplification Little variation High variation High variation Highly optimized High costs Low costs How to learn from different domains? Industrial Transfer Learning 4 Chair of Technologies and Management of Digital Transformation, University of Wuppertal
Industrial Transfer Learning Challenges for Machine Learning in Production Process variations lead to high learning effort for AI e.g. new product, other material, tool change, new machine Product C Product A Product B New Data New Data New Data & Training & Training & Training How to overcome process variations? Industrial Transfer Learning 5 Chair of Technologies and Management of Digital Transformation, University of Wuppertal
Industrial Transfer Learning Transfer Learning – An Emerging Paradigm What is Transfer Learning? Source Tasks Target Task Traditional ML: learning a problem from scratch Transfer Learning: use of existing knowledge [1] Model Model Knowledge Result: faster learning process with less target data “Transfer learning will be the next driver of ML success.” Andrew Ng, NIPS 2016 keynote [1] Pan, Sinno Jialin, and Qiang Yang. "A survey on transfer learning." IEEE Transactions on knowledge and data engineering 22.10 (2010): 1345-1359. Industrial Transfer Learning 6 Chair of Technologies and Management of Digital Transformation, University of Wuppertal
Industrial Transfer Learning Transfer Learning – State of the Art Use Cases of Deep Transfer Learning Self-Driving Cars Robotics Music Computer Vision Natural Language Classification Processing Use of large datasets Use of pretrained Use of simulation Pretraining in Transfer of pattern for classifying music language models for environment to train simulation for recognition (e.g. genre specific NLP tasks artificial intelligence grasping and edges, objects) to manipulation new images Industrial Transfer Learning 7 Chair of Technologies and Management of Digital Transformation, University of Wuppertal
Industrial Transfer Learning Industrial Transfer Learning – A Definition In the field of production, industrial transfer learning refers to machine learning methods and techniques that make use of source data from different production process domains or process variations with the goal to create robust, accurate and data efficient models for a certain target task . Process domain Process variation Real Machine Product Pre-production Material Expert Knowledge Tool Simulation Machine Industrial Transfer Learning 8 Chair of Technologies and Management of Digital Transformation, University of Wuppertal
Industrial Applications Simulation to Reality Transfer for Predictive Quality
Simulation to Reality Transfer for Predictive Quality Predictive Quality in Injection Molding Supporting process designers in the initial set-up of a machine by predicting quality criteria from machine parameters Increasing data efficiency by transfer learning from simulation to real world Conducting design of experiments on real machine and simulation with six parameters Cavity Temperature Cooling Time Melt Temperature Injection Time Quality Holding pressure level (part weight) Holding pressure time Plate Specimen Industrial Transfer Learning 10 Chair of Technologies and Management of Digital Transformation, University of Wuppertal
Simulation to Reality Transfer for Predictive Quality Bridging the Reality Gap Transfer Learning Model Training ▪ Pretraining in simulation ▪ Neural network with two hidden (Cadmould 3D-F) layers with 40 neurons ▪ Finetuning of the network ▪ Activation function: tanh Machine Part weight Parameters Pretraining Finetuning (simulation) (real data) Industrial Transfer Learning 11 Chair of Technologies and Management of Digital Transformation, University of Wuppertal
Simulation to Reality Transfer for Predictive Quality Successful Transfer Reduction of Training Effort Use of simulation data improves prediction models for Transfer Mit TL real process Baseline Ohne TL Improvement in accuracy by factor of 3 0 5000 10000 15000 20000 25000 Number of Training Iterations Reduction of learning effort (iterations) by 80% Increasing Data Efficiency Transfer 1 Performance 0,6 0,2 -0,2 Without Transfer Ohne TL Transfer Mit TL -0,6 Pretrained from -1 simulation 1 10 20 30 40 50 60 Number of Real Experiments Industrial Transfer Learning 12 Chair of Technologies and Management of Digital Transformation, University of Wuppertal
Simulation to Reality Transfer for Predictive Quality Successful Transfer Continuous improvement of model Simulation AI-Model by new simulated experiments Trigger ▪ AI bridges the gap between simulation and Training real manufacturing process ▪ Use for automated design in production line Process ▪ In case of uncertain predictions: − Automatic triggering of new experiments Adjustment in simulation − Transfer of newly gained knowledge to Control real process Industrial Transfer Learning 13 Chair of Technologies and Management of Digital Transformation, University of Wuppertal
Industrial Applications Continual Learning of a Predictive Quality Model
Continual Learning of a Predictive Quality Model Predictive Quality in Injection Molding Cavity Temperature Cooling Time Melt Temperature Predicting quality criteria from machine parameters by means of a neural network Injection Time Quality Holding pressure level (Deformation) Holding pressure time Difference of quality for different products Production of a new product variants Changes in geometry and process behavior ➢ Predictions no longer work ➢ Requires training of a new prediction model Industrial Transfer Learning 15 Chair of Technologies and Management of Digital Transformation, University of Wuppertal
Continual Learning of Predictive Quality Model Use of Previous Knowledge for Transfer Product 3 Product 4 Product 1 Product 2 Transfer Transfer Transfer Learning without forgetting Learning capability increases Amount of data decreases Industrial Transfer Learning 16 Chair of Technologies and Management of Digital Transformation, University of Wuppertal
Continual Learning of Predictive Quality Model Incremental Learning without Forgetting Retuning Finetuning Learning without forgetting process specific product specific … Product 1 Product 2 Product 3 Industrial Transfer Learning 17 Chair of Technologies and Management of Digital Transformation, University of Wuppertal
Continual Learning of Predictive Quality Model Improving Efficiency and Learning Improved Performance Improved Data Efficiency ▪ Continual learning approach keeps ▪ Number of required training data is up performance reduced for every product ▪ Traditional approach becomes worse ▪ Prediction model can generalize with every product better to new parts 80 100 70 # Training Data Performance 60 90 50 40 80 30 20 70 10 Products 1st 2nd 3rd 4th 5th 6th 0 Products 1st 2nd 3rd 4th 5th 6th Continual Learning Continual Learning Learning from Scratch Learning from Scratch Industrial Transfer Learning 18 Chair of Technologies and Management of Digital Transformation, University of Wuppertal
Industrial Applications Sim2Real Transfer for Reinforcement Learning in Robotics
Sim2Real Transfer for Reinforcement Learning in Robotics Reinforcement Learning Automated Trial-and-Error by Learning AI Model ▪ AI agent learns by means of interactions with its environment – Agent observes state – Agent chooses action – Environment issues reward ▪ Actor-critic architecture – Critic: learns the action-value function – Actor: specifies the current policy ▪ Deep Deterministic Policy Gradient (DDPG). – Used for a number of continuous control tasks in simulated environments Industrial Transfer Learning 20 Chair of Technologies and Management of Digital Transformation, University of Wuppertal
Sim2Real Transfer for Reinforcement Learning in Robotics Use of DDPG in the Real World ▪ The wire loop game as an easy-to-control sandbox scenario. – State : camera images, Action : three degrees of freedom (forward, sideways, rotation), Reward : contact between fork and wire image processing (CNN) decision making (FCNN) camera images execution of motion current signal ! High training effort on real industrial robot! Industrial Transfer Learning 21 Chair of Technologies and Management of Digital Transformation, University of Wuppertal
Recommend
More recommend
