CFS Europe

Process optimisation and sustainability pass, first of all, through production data control and understanding. Data decentralisation and heterogeneity make ordinary operations time consuming and potentially corrupt. The challenge was to digitise the whole process and create a system to detect and predict any unusual condition and its causes to optimise performances.  

With over 2000 signals, collected in real time from the plant machineries and other data such as laboratory analysis and maintenance history from different data sources, the first phase of the project was focused on the creation of a data collector via FactorAI Data Management Module.

The system was configured to match the need of the CFS management and enable the complete vision of reason why and time length of downtimes, including:

• real time alerting system (via telegram bot) 
• abstract view on the existing  process  and data sources
• interactive user interface to manage and visualise all data layers

Starting from data of the past three years, FactorAI SPC module was used to build a Multivariate Statistical Process Control, in order to monitor relevant performance indicators, considering selected variables related to each specific KPI (energy and raw material consumption, quality control, production efficiency…).

Besides, by integrating a Principal Component Analysis in the MSPC Model, an anomaly detection system was set to determine drifts of variables impacting KPIs. 

In the end, a Synoptic Map and a root-cause analysis have been integrated to help the end-user to investigate the reasons behind a drift of a KPI, pinpointing variables drifting from normal behaviour. By suggesting actions to take on the plant, the platform works like an interactive operative manual.

Through the FactorAI Predictive AI Module, a Real-time process monitoring and alert platform has been developed. 

To detect the plant’s state, an xgboost machine learning model was trained, using all the DCS tags available and laboratory analysis results. The model achieved high prediction performances (~98%), but using all the variables as input in a production environment is computationally unsustainable. Therefore, by analysing the gain contribution of the most important variables for the prediction, a small subset of data was selected (~15) and a second xgboost model was trained.

The second model achieved an accuracy almost identical to the first one and has been implemented in production. The variable selection step allowed for a noticeable reduction of the prediction times while maintaining the same optimal accuracy performance.