Predictive-maintenance monitoring¶

Predictive maintenance often starts with the same practical need: rank machine states by how unusual their multivariate sensor pattern looks.

Result at a glance¶

The synthetic monitoring example reaches precision and recall around 0.786. It is intentionally less perfect than the simple sensor example, which makes it a better reminder that faults may overlap with normal operating variation.

What the data represent¶

The simulation creates time-like machine states with correlated sensor features and injected degradation/fault periods.

Why this estimator¶

FastMCD or AutoRobustAnomalyDetector are reasonable first choices. The robust distance becomes a health score that can be tracked over time.

Reproduce the result¶

python examples/use_case_maintenance_monitoring.py

Output from the run¶

predictive-maintenance monitoring
precision=0.786, recall=0.786, detected=70
radial_kurtosis=1.607
saved diagnostics to results/use_cases/maintenance

Figures and diagnostics¶

Predictive-maintenance monitoring — distance panel

How to read the result¶

The time profile is the most useful plot. Look for sustained runs above threshold rather than isolated single-point spikes; sustained elevation is usually more actionable for maintenance.

What this does not prove¶

Production monitoring should include temporal smoothing, operating-mode segmentation, and feedback from maintenance events.