Business Context

VoltGrid Manufacturing operates 1,200 industrial machines across 18 plants and wants to predict equipment failures 24 hours in advance so maintenance can be scheduled before unplanned downtime occurs. False negatives are costly because each missed failure can stop a production line, but too many false positives overload the maintenance team.

Dataset

The training data is built from machine-hour snapshots over the last 24 months.

• Size: 420K machine-hour records, 52 features
• Target: Binary label indicating whether the machine fails in the next 24 hours
• Class balance: 2.7% positive, 97.3% negative
• Missing data: ~8% missing in sensor features due to telemetry dropouts; ~3% missing in maintenance fields for newly installed machines

Success Criteria

A good solution should achieve strong ranking quality and support an operating point that catches most failures without overwhelming maintenance capacity. Target at least 0.45 PR-AUC, 0.88 ROC-AUC, and 70% recall with precision above 25% on a held-out time-based test set.

Constraints

• Predictions run hourly in batch for all active machines
• Inference per machine should stay under 50 ms in production
• The maintenance team needs feature-level explanations for flagged machines
• Training data is temporal, so leakage from future information is not acceptable

Deliverables

1. Select a baseline and a primary model for failure prediction, and justify both.
2. Build a leakage-safe preprocessing and training pipeline.
3. Tune the model using time-aware validation.
4. Choose a decision threshold based on maintenance team capacity and failure recall.
5. Report evaluation metrics, feature importance, and key production risks.