UC 5.2 Kappa Sensor
Södra and ABB are collaborating on this use case. They’re working together to enhance its capabilities and find innovative solutions. Their partnership demonstrates a shared commitment to improving processes and driving progress.
The Kappa sensor is direct input to the Advanced Process Control (APC) of the pulping process. The APC controls the quantity of raw material that enters the process as well as other relevant variables, e.g., steam temperature and pressure, chemicals, and machine speed to ensure desired paper quality.
Process Section: Digester
Impact: Impact on product quality (material can become harder) and possible increased material (chemicals) consumptions, throughput due to longer time to digest the raw material.
Desired Outcome: From the soft sensor itself, an estimate of the Kappa value. In addition, information about the quality of the estimation, the health of the sensor, guidance when to retrain the AI model, whether additional model version for different raw material is required, providing an understanding on how the soft sensor works (UX). Possibility to treat the sensor like any other instrument: Compare against lab samples, perform a calibration, possibly using explanations for diagnosing the sensor.
Image Source: Processes 2020, 8, 1231; doi:10.3390/pr8101231
User Interface
Dashboard: Grafana-based and interactive visualization to explain each forecast.
ML Model Summary
Use case: multi-step prediction (60min) to forecast sensor readings that indicate production quality, i.e., Kappa value for pulp & paper.
Explainer Component Overview
- Uncertainty of prediction using quartile envelopes.
- Precision of historic predictions regarding quartile envelopes.
- Self attention-based input signal and time step importance matrix that indicate which input data are essential for the prediction made by the model.
- Time step importance for each individual input signal.
Feedback Component Overview
- Operators labeling phases of the process which operated under sub-optimal conditions
- Using proxy model to simulate the effect of operators’ reprioritizations of feature importance
