Academic Partners: School of Engineering (JTH)
Project duration: 2023-2026
Project managers/Team of researchers:
Tuwe Löfström, Ass. Professor
Rudy Matella, Ass. Professor
Ulf Johansson, Professor
Predictive maintenance using confidence predictors
Predictive maintenance using machine learning for fault diagnosis and remaining useful lifetime estimation is the most effective solution for smart manufacturing. The project explores how confidence predictors can improve decision support in predictive maintenance tasks.
The purpose of the project is to explore how the well-calibrated confidence measures and rich information provided by confidence predictors for classification and regression can improve decision support in predictive maintenance tasks together with our industrial partners Siemens Energy and Jönköping Energi. In regression, these well-calibrated measures take the form of prediction intervals or cumulative distribution functions, and in classification label sets or probability intervals. Typical predictive maintenance tasks involve fault detection and remaining useful lifetime estimation, making it a perfect fit for tools such as conformal anomaly detection and martingales for concept drift.
Importance of the project
Trustworthy machine learning requires algorithmic confidence, i.e., the models must be able to communicate their confidence in every prediction. Correctly estimating the remaining useful lifetime of critical infrastructure in manufacturing processes have the potential to decrease cost substantially. Detecting faults before they have happened or have caused damage can help avoid unnecessary downtime. We argue that the Prediciton with Confidence framework have the necessary tools needed to provide algorithmic confidence to typical tasks within the field of predictive maintenance.
The expected results include both scientific contributions related to algorithmic confidence and solutions to industrial challenges. An initial industrial challenge to address at Jönköping Energi is to utilize confidence predictors to better understand the cause-and-effect leading up to clogging of the heat and power plant at Torsvik, and to accurately predict when to plan for maintenance to prevent clogging. At Siemens Energy, a recurring problem when producing large heat-resistant disks is that the geometry of the discs ends up incorrect. Creating accurate fault detection conveying confidence estimates and distinguishing among material-dependent, machine-dependent and external factors causing these geometric faults will be another initial industrial challenge to address.
Interested in finding out more?