September 5, 2017
Predicting Product Quality as Early as Possible in the Production Process
Speaker: Dr. Gabriel Fricout, Arcelor Mittal
Data mining processes have proven to be very valuable for addressing industrial issues such as understanding defect crisis. In classical data mining procedures, only “single value” variables are considered, meaning that one individual, in the steel industry typi- cally one coil, is characterized by average values of many process parameters (composition, temperature, speed, strengths, tractions, composition, etc.), which will be used to predict, forecast, or estimate unknown properties about the product, such as the probability of defect occurrence for instance. However, in many situations, the available information is much wider: Many sensors continuously register information about the product and the process. This talk focuses on presenting both methodology and tools to perform data mining using time series from the steel manufacturing pro- cess sensors. The methodology involves several steps of data preparation, statistical modelling and training (based on shapelet and deep learning methods), performance evaluations, and knowledge capitalization. Results are illustrated on real data from the steel production process trying in particular to forecast as early as possible the risk of product non-quality. Tools involved and developed are technological choices based on various available pieces of software (MongoDB, RapidMiner, Kasem). All the develop- ments are conducted within the PRESED RFCS framework.
Dr. Gabriel Fricout, Arcelor Mittal
Head of Surface Properties, Data and Signal Processing Gabriel Fricout holds a PhD in mathematical morphology.
Since 2013 he leads the “surface properties, data and signal processing” (SPID) team within ArcelorMittal’s Measure- ment and Control R&D cluster. The main focus of this 10 persons R&D team, including three PhD students, involves on-line surface characterization through the deployment of industrial innovative sensors, real-time processing of the data coming out of these sensors for process and product quality improvement and data mining. In the framework of his research activities, Gabriel Fricout has supervised 6 PhDs covering various areas of physics, including fluid mechanics and spectroscopy, or computer sciences, including reinforcement learning, deep learning and data mining. Gabriel Fricout also is the coordinator of the EU-sponsored RFCS project PRESED (Predictive Sensor Data Mining) dedicated to time series data mining and member of the Intelligent Integrating Manufacturing (ESTEP / EUROFER). PRESED is a project in- volving 6 partners, including RapidMiner, dedicated to data mining on time series.