How ML Product Development for Manufacturing Works at EuroChem

Vyacheslav Kozitsyn25 May 2024

Problem

EuroChem's AI department has more than 20 data scientists running over 10 industrial projects a year, and each project has about 3 scaling objects. Without one methodology, every solution has its own code structure and cannot be taken into support without consulting the author. Deployment waited weeks for overloaded developers, and data exports waited weeks for on-site staff. Plant data adds its own traps: missing values, outliers, shifting regimes, non-stationarity and mixed-up sensors.

Example from the article: a recommender for process control operators that suggests parameter values to raise the output of a useful product by 5% against the state before the system.

Solution in brief

Every project follows one checklist of stages with fixed durations: task study, data validation and EDA, modeling, commissioning. A plant is described not by one model but by a set of models built from a decomposition tree of objectives. Solutions live in a standard repository and are deployed through CI/CD.

Questions this case answers
  • How to organize many separate ML projects for process optimization so that each solution is supportable by any team, from data validation to commissioning
  • How to model a multi-unit continuous process as a set of forward and reverse models when sensor data is noisy and regimes shift
  • How to organize an ML project for process optimization from data validation to commissioning
  • How to validate sensor data before modeling an industrial process

Tools and methods

  • Forward and reverse tasks — a regression model forecasts output, then an optimizer finds lever values that raise it
  • Tag validation and ordered preprocessing — remove broken tags and bad regimes before any modeling
  • Decomposition tree plus an aggregated model — split a plant into stage-level models and prioritize them
  • Standard repository, MLflow, CI/CD, Grafana alerts — one structure for training, inference and monitoring
Continue reading — it's free

Google or a one-time email code. No password.