Analytica Data Science SolutionsContact
Services
Work
Products
Company
ESC

↑↓ to move↵ to open

← All work

Case study

Predicting equipment failure before it stops the site

Unscheduled downtime is the most expensive thing that happens on a mine site. The operator wanted to stop reacting to failures and start scheduling around them.

Client
A mining and heavy-equipment operator
Sector
Mining · Industrial
Role
Direct engagement
Period
Not disclosed
Result
−23% unscheduled downtime, client-reported
unscheduled downtime, client-reported
−23%unscheduled downtime, client-reported
maintenance cost, client-reported
−12%maintenance cost, client-reported

The problem

Heavy equipment fails, and when it fails unexpectedly the cost isn't the part — it's the shift that stops. The operator was maintaining reactively and wanted to know which components were approaching failure while there was still time to schedule around them.

What we did

Two complementary approaches. On components with enough historical failure data, we modeled remaining useful life directly — how much service life is left in this specific part, given how it has been used.

On components without that history, supervised learning has nothing to learn from. So we used unsupervised methods to estimate stress levels from operating data instead, which flags parts behaving unlike their peers without needing to have seen them fail.

The result

The operator reported unscheduled downtime falling by more than 23%, and maintenance costs by more than 12%, once preventive scheduling was in place.

The second-order effect the operator valued more: work could be planned. Knowing a component has weeks rather than days left is what lets a maintenance team order parts and book a window instead of scrambling.

Predicted remaining useful life against a failure thresholdComponent health declines with operating hours. Telemetry ends partway along; the model continues the curve into a prediction interval that widens with distance. The interval crosses the failure threshold at a known point, and the shaded band before that crossing is the window in which maintenance can be planned rather than scrambled — parts ordered, a slot booked, the machine taken down deliberately.FAILURE THRESHOLDLAST READINGPLAN HERESCRAMBLE HEREOPERATING HOURS →100%0%COMPONENT HEALTH
A model that tells you a bearing has failed is a log entry. A model that tells you it has roughly three weeks left lets the team order the part, book a window, and take the machine down deliberately — that gap between the two labels is where the client-reported 23% reduction in downtime came from.

Discuss a similar problem

If one of these engagements resembles a problem you're facing, we can walk through how it was built and what it would take in your environment.

Start a conversation