Business Context

NorthRiver Manufacturing operates 1,200 industrial pumps across 18 plants. Unplanned pump failures cause expensive production downtime, so the maintenance team wants a model that predicts whether a machine will fail within the next 7 days using recent telemetry and maintenance history.

Dataset

You are given one row per machine-day, built from rolling aggregates over raw sensor streams.

• Size: 420K machine-days across 24 months, 46 features
• Target: Binary label indicating whether the machine fails within the next 7 days
• Class balance: Highly imbalanced — 2.9% positive, 97.1% negative
• Missing data: ~8% missing in sensor features due to telemetry dropouts; 12% missing in maintenance logs for newly onboarded plants

Success Criteria

A good solution should identify at least 75% of upcoming failures while keeping precision above 35%, and provide ranked risk scores that maintenance planners can use for daily scheduling.

Constraints

• Daily batch scoring must finish in under 10 minutes for all active machines
• The solution must be explainable enough for reliability engineers to trust flagged machines
• False negatives are costly, but too many false positives will overwhelm maintenance capacity

Deliverables

1. Build a supervised learning model to predict 7-day failure risk.
2. Explain feature engineering, leakage prevention, and model choice.
3. Define a time-based validation strategy and threshold selection approach.
4. Report evaluation metrics appropriate for imbalanced classification.
5. Describe how you would deploy and retrain the model in production.