Digital inspector - Federal Fleet Learning

Challenge

Industrial machines generate growing volumes of sensor data that can support predictive maintenance and usage monitoring. However, turning this data into reliable insights across an entire fleet remains challenging.

Failures are often rare, which means learning from a single machine can take too long. At the same time, machines within a fleet can have different configurations and operate under different environmental and operational conditions. Standard machine learning models struggle to transfer knowledge effectively between these different contexts.

This can lead to a difficult trade-off: highly specialised models result in many models to maintain, while overly generic models can require excessive resources. Collecting and correctly labelling data presents an additional challenge, particularly when sensor coverage or contextual information is incomplete.

Project Goals

With this project (Digital_Inspector_IRVA), we aim to develop a scalable fleet-learning toolbox for predictive maintenance and usage monitoring, enabling companies to learn across machines instead of relying solely on years of data from individual assets.

The project will develop fleet-oriented normal behaviour modelling and anomaly detection that can transfer knowledge between heterogeneous machines while adapting to differences in configuration and operating conditions. Continuous model updating, uncertainty assessment and automated data selection will help maintain model confidence as conditions change.

A second focus is scalable data collection and automatic labelling when information is incomplete, including non-invasive vibration monitoring using fibre-based sensing. The resulting modular, cloud-native building blocks will allow companies to select and integrate the capabilities relevant to their fleet and use case. For the defined project use cases, the project targets precision and recall of at least 80% for failure prediction and data labelling.