EUROPEAN STEEL TECHNOLOGY AND APPLICATION DAYS 2017 - ESTAD 2017

Data mining process has proved to be very valuable for addressing industrial issues such as understanding defect crisis. However, for the methodology to be successful, several key criteria have to be fulfilled: * Data has to be relevant and well prepared * Rigorist mathematical, statistical and modeling methodologies have to be used to assess result robustness. * As many process knowledge as possible has to be integrated in the analysis and in the result interpretation for more relevancy In classical data-mining procedure, only "single value" variables are considered, meaning that one individual, typically one coil for the steel industry, is characterized by average value of many process parameters (composition, temperature, speed, strengths, tractions, composition ...), which will be used to predict, forecast, 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 since many sensor continuously register information about the product and the process. Theoretical tools to conduct data-mining studies with such high dimensional "time-series" data are progressively developing, following the "big data" trend of the last years. However, many open questions remains for a concrete industrial application like defining accurate statistical defect models for the steel industry. In the PRESED RFCS (PRedictive Sensor Data-mining), the overall methodology to build a statistical defect model from time-series data bases is investigated, including the development of new theoretical methods based on "shapelet" theory, but also the use of specific ontology tools to integrate process knowledge in the model and IT infrastructure (data-base, analytics server) to get practical and usable results from realistic industrial data sets. The paper will focus on presenting this overall methodology as well as practical example of use like sliver defect modeling or mechanical properties scattering.