CUSTOMER STORY | AEROSPACE
Safran anticipates machine failures 30 minutes ahead, with a 94% recall rate
Safran wanted to predict breakdowns on the machines that manufacture aircraft engine blades around the clock. TIMi built predictive models from existing production data, without installing any new sensor.
Failure recall rate
94%
Advance warning
30 min
First model deployed
Under 1 week
OVERVIEW
The challenge.
On a 24/7 production line, one failing machine stops everything. The aim was to catch rare failure signals in a very large volume of normal data.
One machine stops the line
The 7-axis machines that make aircraft engine blades run 24/7. When one of them fails, the entire production line comes to a halt.
10 billion rows, few failures
10 production machines, each generating over a billion rows. Less than 0.5% of these rows correspond to a breakdown.
No additional sensor
The signals had to be captured from existing data, without installing any new sensor on the machines.
SOLUTION
What was done.
TIMi went on site to deploy a first predictive model, then the project continued to push the models further.
Deploy
A first predictive model was deployed on site in under 5 days, relying purely on existing data.
Refine
The project continued with the Safran team, who pushed the models further to extend the warning time and raise the recall rate.
Understand
As a by-product, the entire 20-step manufacturing sequence was reconstructed to pinpoint the exact mechanical command causing factory-wide vibrations.
RESULTS
What changed.
Technicians now intervene proactively, before a failure stops the line.
First model
An 85% recall rate, with failures anticipated 15 minutes early.
After refinement
Anticipation time doubled to 30 minutes, and the recall rate reached 94%.
Operations
Technicians now intervene proactively instead of reacting to a stopped line.
Testimonials
Their feedback
“A first attempt based on Python proved poorly suited: analyzing a single file out of millions required nearly two weeks of computation. With TIMi and a correctly sized server, the whole pipeline could be applied to the entire database in only one week.”
Hugo D. – Predictive maintenance intern, Safran
SEE FOR YOURSELF
Test it on your own data.
Run TIMi on your own data with a free trial, or talk to us about your case.
