INDUSTRY & MAINTENANCE
Predictive maintenance: fix it before it stops
Predictive maintenance with TIMi anticipates the failure of an equipment from the operating data you already collect, so an intervention can be planned before an unplanned stop, and shows the factors behind each alert.
Risk per machine
Each equipment gets a risk level at a horizon you define.
Works with existing data
Controllers, logs and production databases are often enough to start.
Explained alerts
Each alert tells the team what to check.
HOW IT WORKS
How predictive maintenance works.
Three steps, in one engine, on a standard server with no cluster.
1. Prepare the data
Anatella gathers the operating data (sensors, operating logs, maintenance history) and the history of past failures, then structures them into variables.
2. Build the model
Modeler compares the signals that came before known failures with the ones observed now, and estimates a risk of failure at a given horizon, even when the failure history is limited.
3. Alert and explain
Each equipment gets a risk level with the factors behind it, such as a drift in an operating parameter or a combination of signals already seen before a past failure.
Anatella
ModelerGETTING IT RIGHT
Choices that shape a maintenance model.
Failures are rare, and that shapes the approach.
Work with rare failures
Failures are a tiny share of the data. At Safran, fewer than 0.5% of 10 billion rows corresponded to a breakdown. Modeler is designed to extract a usable signal from a limited failure history, using all the operating variables available.
Start with the data you have
A model can often be built from data already collected by existing systems, before considering extra sensors. At Safran, the first model was deployed within a week with no new sensor.
Plan for changing conditions
A change of operating conditions, such as a new pace or a new environment, can change the warning signals. Modeler’s Drift Report shows when a model needs an update. At SANY, models built in summer decayed in winter, and the Drift Report made the shift visible.
EXPLAINABILITY
An alert says it will fail. The factors say what to check.
An alert alone says an equipment is at risk, not why. A white-box model shows the factors behind the alert for this machine, so maintenance teams can prepare the right intervention (part to replace, type of check) instead of starting from a generic signal.
For maintenance teams
An intervention can be planned and equipped in advance, knowing the probable cause of the risk.
For trust in the system
Technical teams adopt more easily a system whose logic they can check than an alert they cannot question.
For model owners
Modeler’s built-in Drift Report shows when a model loses accuracy, so you update it when needed instead of on a fixed calendar.
DATA
What data a maintenance model uses.
Anatella connects to more than 80 sources, so this data can be gathered without prior custom development.
Sensor and operating data
Temperature, vibration, pressure, depending on the equipment.
Logs and maintenance history
Operating logs and maintenance already carried out.
Past failures
Even in limited number.
Operating conditions
Load, environment, pace.
PROOF
Predictive maintenance in production.
Three industrial deployments, in aerospace, mining and oil.
94% of failures anticipated 30 minutes ahead
Machines that make aircraft engine blades 24/7. 10 billion rows analyzed on a standard server, with no new sensor.
Engine failures predicted within the hour
AUC above 98% with 355 correlated variables and only 107 historical failures.
€19.6 million of repair costs avoided a year
Failure models on 5,000 wells and 500 sensors also added €1.3 million of revenue a year.
FAQ
Common questions.
Do we need many past failures to train a model?
Not necessarily. Modeler is designed to extract a usable signal from a limited number of failures, using all the operating variables available rather than the volume of incidents alone. At SANY, the model was trained on only 107 historical failures.
Do we need to install new sensors?
Not necessarily. A model can often be built from the data already collected by existing systems, before considering extra sensors if a missing signal is identified.
How long does it take to deploy a first model?
Once the operating data is gathered and accessible, a first model can be built quickly. At Safran, a first model was deployed on site within a week.
Does the model need retraining if operating conditions change?
Yes, a change of conditions can change the warning signals. Modeler’s Drift Report shows when an update is needed.
RELATED USE CASES
Keep exploring.
Equipment Failure Prediction
Estimate the probability that an equipment fails at a given horizon, and why.
Equipment Downtime Prediction
Estimate how long a stop will last and what it will cost in output.
Model Drift Monitoring
See when a predictive model loses accuracy and needs an update.
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.
