Anticipating equipment failure.
WHAT IT PREDICTS
That equipment will fail within a window, or its remaining useful life.
WHAT DATA IS REQUIRED
Sensor readings over time Maintenance records Failure events, with dates Operating conditions
WHAT THE CENTRAL DIFFICULTY IS
Failures are rare, and well-maintained equipment produces few.
WHAT THAT MEANS
There may be too few examples to learn from.
WHAT TO DO ABOUT IT
Consider anomaly detection instead of supervised prediction.
WHY
It learns normal behaviour, which is abundant.
WHAT TO BE CAREFUL WITH
Maintenance records that are incomplete or entered late Sensors replaced or recalibrated, changing the signal Failures prevented by maintenance, which never appear as failures
WHAT THAT LAST POINT MEANS
Successful maintenance removes the evidence you need.
WHAT THE HORIZON MUST ALLOW
Time to schedule and perform the work.
WHAT TO MEASURE
Failures prevented Unnecessary maintenance caused Downtime avoided
WHY THE SECOND ONE
Excessive false alarms make the system ignored.
WHAT MATTERS LOCALLY
Power fluctuation as both a failure cause and a signal.
