Business Context

NorthRiver Manufacturing operates 1,200 industrial pumps across 18 plants and wants to predict whether a machine will fail within the next 7 days so maintenance can be scheduled before unplanned downtime occurs. False negatives are costly because a missed failure can stop a production line, but too many false positives create unnecessary maintenance work.

Dataset

You are given a historical predictive maintenance dataset built from sensor readings, maintenance logs, and machine metadata.

• Size: 96K machine-day records across 24 months, 44 features
• Target: Binary label indicating failure within the next 7 days
• Class balance: 6.5% positive, 93.5% negative
• Missing data: ~12% missing in some sensor windows due to telemetry dropouts; 4% missing in maintenance fields for newly onboarded assets

Success Criteria

A strong solution should achieve recall >= 0.80 on failure events while keeping precision >= 0.35 and PR-AUC >= 0.40 on a held-out time-based test set. The candidate should explain how model complexity affects underfitting vs overfitting in this maintenance setting.

Constraints

• Predictions run every 6 hours in batch for all active machines
• Maintenance managers need a reasonably interpretable model and feature importance
• Training data is temporal, so leakage must be avoided
• Inference should remain under 200 ms per machine in offline batch scoring

Deliverables

1. Build and compare at least one low-variance model and one higher-capacity model
2. Explain the bias-variance trade-off for predictive maintenance and how it appears in train vs validation performance
3. Design a leakage-safe validation strategy for temporal data
4. Recommend a final model and decision threshold for operations use
5. Describe how you would monitor drift and retrain the model in production