Time Series and Predictive Maintenance

Inspect timestamps, time zones and irregular intervals.

Evidence: A cleaned timeline retains an audit of dropped or repaired points.

Aggregate sensor streams without filling across invalid gaps.

Evidence: A gap policy passes supplied edge cases.

Split by time and respect asset boundaries where required.

Evidence: A diagram shows prediction horizon and feature availability.

Build last-value and seasonal baselines.

Evidence: Forecast error is reported over a held-out future window.

Create lag and rolling features using past values only.

Evidence: Unit tests verify no future observations enter any feature.

Compare a statistical or regression model with the baseline.

Evidence: Rolling-origin evaluation supports the comparison.

Compare rule limits and an unsupervised detector.

Evidence: Anomaly scores and review assumptions are explicitly documented.

Estimate event risk where reliable labels exist.

Evidence: A label definition distinguishes censored data and confirmed failures.

Tune persistence and suppression rules for noisy signals.

Evidence: An alert simulation reports missed events and alerts per asset.

Combine signals and maintenance notes without claiming causation.

Evidence: A diagnostic narrative separates evidence from hypotheses.

Replay events through an alert pipeline.

Evidence: Late events and reconnects do not create duplicate alerts.

Define triage, escalation and feedback capture.

Evidence: A runbook links alert severity to an accountable human response.