To make the most of your techUK website experience, please login or register for your free account here.
AtkinsRéalis is leading a project that is a proof of concept for predictive maintenance at a modern Nuclear Power Station. The aim is to demonstrate that machine learning can predict and prevent component failures, reducing plant outages and preventing downtime.
Advanced predictive maintenance can improve policies and reduce cost. Understanding the probability of failure has the potential to transform the organisation’s maintenance strategy, moving it from fixed interval inspections to targeted component inspections focused on the equipment with the highest probability of failure. At the same time, we can reduce the frequency of inspections on components that have low predicted probability to fail.
By using supervised and unsupervised machine learning to map historical data of failures to historical sensor readings, we are extracting patterns that could predict future failures. To do this, we are leveraging a range of different approaches to anomaly detection techniques, ranging from clustering-based approaches to deep learning algorithms, to train and validate our failure-prediction models. These are implemented across several different platforms.
One of the primary challenges of the project has been deciding which plant systems to target, particularly as knowledge has been segmented and distributed across multiple stakeholders. By the project’s conclusion we hope to have:
• A working prototype that could form the basis of a pilot deployment
• Tangible evidence that predictive models and machine learning can be used to predict and prevent failures
• A way to improve predictability of failures and prevent system downtime and costly outages
This model can then be used to equip the newest generation of UK nuclear plants with operational predictive models from day one sharing knowledge and operational experience across the Nuclear and Power industry.
