Symposium on the occasion of the 5th anniversary of the Institute for Carbon Composites. Dr.-Ing. Christian Sorg - BMW AG
Agenda. (1) Introduction Requirements to high-volume Composite Manufacturing Challenges (2) Knowledge Discovery in Databases (3) Data Mining System Modeling of the process chain Concept and setup of the data mining system Capabilities of a data mining environment (4) Examples: Reporting example: causes of defects Analytics example: CHAID decision tree Predictive analytics example: failure detection system (5) Further possibilities: support of the ramp-up management (6) Benefits and conclusion 2
Quantity Requirements to high-volume composite manufacturing. Requirements to highvolume manufacturing Automotive: high-volume Reduction of material costs Reduction of cycle-times Aerospace: low-volume Machinery Capable Process Control Systems Motorsports: manufacture 3
Functional Process Model Challenges to high-volume Composite Manufacturing. Problem Consequence Needs Highly interacting processing parameters Sequenced process chain (RTM- Process) carries defect parts through the process Sensitive process spontaneously occurring failures Extensive quality controls after every process step (early failure detection necessary) high costs Capable process control system Relation between processing parameters and final part properties not entirely resolved Time consuming fault-removal process high idle time high costs Process knowledge and transparency 4
Handling of interactions. Schematic view of a 2-parametric interaction: Vektor transformation Not separable by tollerated single parameters A A t B z B A t B Example: Support Vector Machine (SVM) Examplepart: ( ): parameter A parameter B reworking issue 5
Knowledge Discovery in Databases and Data Mining. Traditional approach: manual acquisition, analysis and interpretation of data by experts Problems: rising complexity of manufacturing processes + dramatically increasing volume of data to handle Traditional approaches are stretched to their limits Problem-Solving-Approach: Knowledge Discovery in Databases KDD is a structured process with the aim of identifying valid, new, useful and interpretable information within huge data collections Data Mining Data mining is a part of the KDD process in which methods from different fields (statistics, artificial intelligence, machine learning, ) are applied on the current data Data Source Integration of physical knowledge Selected Data Transformed Data Data Mining Preprocessed Data Patterns interpretation iterative process (expert knowledge necessary) Knowledge Fayyad 1996: Knowledge Discovery in Databases (KDD) 6
BMW J Modeling of the process chain. Preforms Disturbance Variables: Material Measurement Errors Control Variable Tolerances Temperature, applied Pressure Environment Humidity, Temperature Input Variables: Material Resin Hardener Releasing Agent Disturbance Variables: Material Matrix, Fiber Control Variable Tolerances Temperature, applied Pressure Environment Humidity, Temperature CFRP- Components Stack BLACK-BOX Preforming process fpreform( ) Resin System Preform BLACK-BOX RTM-Process frtm( ) RTM-Part Moses Lake, USA Input Variables: Material Layers Fabrics Type Areal-Weight Non-Crimp-Fabrics Type Areal-Weight Binder VA-Mesh Core Control Variables: Temperatures Heating System Tool-Temperature Pressure Stamp Pressure Time Cycle Time Heating System Cycle Time Preform Press Target Values: Properties Geometry Weight Stability Quality Issues Type of Defect Location of Defect Process Issues Cycle Time Material Requirements Waste Defect Rate Control Variables: Temperatures Resin / Hardener Tool-Temperature Pressure Stamp Pressure Injection Pressure Time Injection Time Demoulding Time Target Values: Properties Fiber-Volume Fraction Geometry Weight Quality Issues Type of Defect Location of Defect Process Issues Cycle Time Material Requirements Waste Defect Rate 7
Concept and setup of the data mining system. Preform Quality RTM Quality Finishing Process gate gate Quality gate 1 2 3 1 2 3 1 2 3 Collection of process data Collection of quality data Collection of process data Collection of quality data Collection of process data Collection of quality data Data acquisition Process flow (product) Data Mining Data flow Movement of a product between single processing steps Data acquisition in local environment Data collection (Data-Warehouse) 8
Capabilities of a data mining environment. Manufacturing process Reporting Visualization of all relevant information Continuous transparency Data (Data-Warehouse) Analytics Evaluation of the process data Using methods from the fields of statistics, and data mining Continuous Improvement Predictive Analytics Prediction of failures and reworking issues using trained data mining models 9
Reporting example: causes of defects. Browser based reporting tool with combined information of process and quality data Type 1 Type 2 Type 3 10
Analytics example: CHAID-Decision tree. Automatically generated decision tree Reworking issue: surface ripples at the BMW M3 Coupé CFRP roof 11
Predictive analytics example: failure detection system. Integrated Quality System for the detection of failures and reworking issues. Advantages: Online monitoring of all processing parameters Visualization of the predicted results support of the machine operator Early detection of defect parts allow early removal from the supply chain increase of the netbenefit Preforming Process RTM-Process Resin- System Stack Preform Part 12
Further possibilities: Support of the ramp-up management Faster identification of the optimal process-window in new processes by the use of data mining ( t ). capable controlled t capable not controlled not capable not controlled 13
Benefits and conclusion. Reporting: Analytics: Predictive Analytics: Increased availibility of all relevant data accelerated visualisation and analysis Support of the problem solving process Faster diagnosis minimization of idle times extended and automated analytical methods by the use of data mining quantifiable influence of control variable to dependent variable Description and handling of multidimensional interactions possible Promising results from a pilot study at the M3 Coupé CFRP-roof Early removal of scrap parts from the process chain to increase cost efficiency Reduced test cycles of the parts Conclusion: Data mining offers a variety of new possibilities to the high-volume manufacturing of composite parts and is able to contribute to the industrialization of the carbon-fiber in automotive industry. 14
Thank you for you attention. Kontakt: Dr.-Ing. Christian Sorg E-Mail: christian.sorg@bmw.de Mobil: +49 151 6022 8310 15