Business Context

AeroSight monitors telemetry from 1,200 commercial aircraft and wants a model for engine health analytics. Depending on the downstream use case, the team may need either a discrete operational decision or a continuous estimate, so you must decide when classification is appropriate versus regression.

Dataset

You are given historical flight-level and maintenance data collected over 24 months.

• Classification target: maintenance_required_7d (1 if the engine required unscheduled maintenance within 7 days, else 0)

• Regression target: remaining_cycles_to_service (continuous estimate of cycles until next required service)

• Size: 96K flight records, 38 features

• Class balance: 11% positive for maintenance_required_7d

• Missing data: 8% missing in some sensor aggregates due to intermittent telemetry dropouts; 3% missing in maintenance logs

Success Criteria

A good solution should clearly justify when to frame the problem as classification versus regression, build one model for each target, and compare them using appropriate metrics. For classification, target F1 >= 0.68 and recall >= 0.75 on the positive class. For regression, target MAE <= 18 cycles.

Constraints

• Predictions must run in under 100 ms per flight record in batch scoring
• Maintenance planners need interpretable drivers of predictions
• Avoid temporal leakage from future maintenance events

Deliverables

1. Explain when classification is the correct framing and when regression is the correct framing for this aerospace problem.
2. Build a classification pipeline for maintenance_required_7d.
3. Build a regression pipeline for remaining_cycles_to_service.
4. Compare evaluation metrics, error tradeoffs, and operational implications.
5. Recommend which formulation you would deploy for two use cases: maintenance triage and service interval planning.