Original practice scenario: For a hypothetical energy-asset monitoring dashboard, sensor messages arrive late and some devices stop reporting. Design features that distinguish a calm process from missing evidence.
Read model answer +
Keep observation value, observation time and data availability separate. A flat line constructed by forward filling is not evidence that equipment remained unchanged.
Work through it
Retain source timestamps, arrival timestamps, sequence numbers and quality flags; measure lateness and the duration since the last valid observation.
Define bounded resampling and imputation rules by sensor meaning, adding explicit missingness and stale-data indicators to features and displays.
Evaluate alerts against labeled historical or synthetic sequences containing outages, drift and late backfills, using event-time windows that reproduce online visibility.
Tradeoffs to discuss
Imputation can stabilize downstream models but obscures outages; dropping incomplete windows avoids invented values while reducing coverage.
Common pitfalls
Do not use a late measurement in a historical prediction made before it arrived.
Track sensor replacement and calibration changes so a device reset is not mistaken for a process anomaly.
Original practice scenario: A hypothetical industrial maintenance team can inspect only a fraction of daily alerts. How would you evaluate a model that ranks assets for review?
Read model answer +
Optimize for a constrained review decision, not overall classification accuracy. The meaning of a positive outcome and the delay before confirmation must be established with domain reviewers.
Work through it
Define an asset-time prediction unit and a future outcome window, excluding observations collected after an inspection or repair decision.
Compare the model with simple threshold and age-based baselines using precision or recall at the actual inspection budget and time-separated validation.
Review missed events, false alerts and coverage by equipment family; retain a human escalation route for uncertain or out-of-distribution cases.
Tradeoffs to discuss
A high-recall ranking can consume scarce inspections; a high-precision ranking may miss uncommon but consequential failure modes.
Common pitfalls
Historical inspections are selective, so unlabeled assets are not automatically healthy negatives.
Never describe a ranking experiment as an automated safety decision or a validated production maintenance policy.
