In the first SMS Industry Data Challenge in 2017, participants were given a real-world dataset with the goal to predict defects in continuous steel casting based on time series of very high dimensional sensor data and process parameters. Participants could thereby contribute to the reduction of waste in the resource-intensive steel industry.
My winning submission in the challenge was based on structured prediction with grid-shaped Max-Margin-Markov networks. The approach was inspired by my previous experience on sequence modelling in business process management and natural language processing.
The project was followed by a collaboration with SMS to improve the prototype and prepare it for deployment in the steel plant.