CAISR+: Predictive maintenance and machine learning

The CAISR+ project was about using recent machine learning methods and working with real industrial data (streaming data, onboard data, or data in databases) related to predictive maintenance. There were two main application tracks:

• usage characterization through machine activity recognition (forklift trucks and heavy-duty trucks)
• modelling survivability of vehicle subsystems (turbochargers, coolant pumps, electric drive batteries).

In addition, CAISR+ included participating in international conferences and organising thematic workshops to establish CAISR internationally as an actor in this field. CAISR+ also included advanced courses on the topic directed at professionals.

Subprojects in CAISR+ were related to predicting turbocharger failures, modelling coolant pump failures, detecting “no fault found”, transfer learning from high-frequency to low-frequency databases, machine activity recognition from CAN data, survivability for electromobility batteries, modelling bus usage and battery health. The machine learning topics covered here were transfer learning, self-supervised feature learning, domain adaptation, variational autoencoders, meta learning and explainability.

The CAISR+ project had an industrial advisory board to keep the research relevant for the industrial partners and to point at important industrial developments. In 2020–2022, our work on machine learning for predictive maintenance resulted in 19 journal papers and 14 conference papers. There was also an industrial demonstrator and the transfer of code from academia to industry.