Improving Predictive Maintenance
Machine Learning –
a new superpower for your maintenance team.
Our client generates renewable energy from seven hydro power sites. They wanted to build predictive models to determine hydro generator component failure risk. The goal was to put in place improvements to preemptive maintenance operations and system monitoring.
Initial focus was on generators, turbines and transformers.
Cutting edge deep learning and survival analysis models were built using various operational and maintenance data sources. Time series (by the minute) SCADA data was augmented with maintenance data to create machine generated labels.
An abstract model of the generator systems was then built that could be linked to operational and sensor data and tuned with the client’s risk assessment for each component.
This showed graphically what was statistically “abnormal”, the root cause and how critical it was. It also allowed you to see patterns for the very first time, for example an individual generator “driving style” and bearing stress.
Earlier prediction x5
The SCADA trained models identified that the generator had moved into a damaged state eight power cycles earlier than had previously been observed by the maintenance team. This was a 5-times boost in predictive performance above the baseline.
Key sensor metrics reduced, 1000’s to 10
Using a finite state machine, sensor metrics were reduced from thousands to ten specific ones that provided the best predictive performance to augment existing plant alerts.
Maintenance activity optimized
By comparing actual periodic maintenance patterns with actual faults unnecessary maintenance was reduced and with new models in place, efficient preventative efforts more accurately targeted.
Contact us at email@example.com if you would like a free machine learning value assessment for your team.