Farm A · Portugal (onshore) Farm B · Germany (offshore) Farm C · Germany (offshore)

Competition Results

6 notebooks · 3 farms · 89 turbine-years · 14 models

Best CARE Score
0.319
AE Thermal · Farm C
Best Gearbox CARE
0.253
Hotelling T² · Farm C
Earliest Detection
99.7%
CUSUM Earliness · Farm C
FP Reduction
93%
Ensemble suppression · Farm B
Models
14
Across 6 notebooks
Events
95
Train + prediction sets
CARE Score Leaderboard — All Models

CARE Score Ranking — All Models & Farms

CARE Components — Best Model per Farm

Anomaly Constellation — Every Fault Event as a Star

Severity vs Earliness — All Fault Events Across All Farms — Size = Persistence · Click any star for details

#FarmNotebookModelCoverageAccuracyReliabilityEarlinessCARE

CARE Score Race

Watch all 14 models compete as data is revealed — animated bar chart race

Models ranked by CARE Score — Bars animate to final values

Speed:
Press Play to start the race

CARE Score Distribution by Farm

Model Family Comparison — Average CARE Across Farms

Gearbox & Bearing Health

Notebooks 1a · 1b — Unsupervised detection, CUSUM, per-turbine thresholds

CUSUM Control Charts — Real Detection Results

CUSUM Control Charts — Gearbox Primary Sensor per Farm (from Notebook 1a)

CUSUM charts
Isolation Forest — Anomaly Score Distribution

IF Score Distribution — Normal vs Anomaly — Good separation = distinct peaks (from Notebook 1a)

IF score distribution
CARE Score Components — All Farms & Models

CARE Components per Farm — Coverage × Accuracy × Reliability × Earliness (from Notebook 1a)

CARE components
Per-Turbine Baseline Temperatures — Why Fleet-Wide Thresholds Fail

Baseline Temperature Range per Farm — Min to Max across turbines

CARE-Optimised vs Baseline Threshold

SAX Discord Detection — Unusual Time-Series Subsequences

SAX Discord Detection — Fault-precursor shape patterns vs Normal events

SAX discords
Power Curve Anomaly — Wind Speed vs Generator RPM

Power Curve: Normal vs Anomaly Events — Deviation from expected turbine performance

Power curve
Severity Index — Top Gearbox Events

Severity vs Earliness — Gearbox Anomaly Events

#FarmTurbineEventSeverityPersist (h)EarlinessRisk

Explainability & Maintenance Reduction

Notebook 1c — Label propagation · SHAP · XGBoost vs Gradient Boosting

Label Propagation — Coarse vs Pseudo-Labels

Pseudo-labels assign anomaly only to rows where sensors actually deviate — more precise than whole-event labels (from Notebook 1c)

Label propagation
SHAP Feature Importance — Global (all farms combined)

SHAP Global Feature Importance — Mean |SHAP value| across test set

SHAP global

SHAP Beeswarm — Feature Direction & Magnitude (global)

SHAP beeswarm global
SHAP Beeswarm — Per Farm (XGBoost with Pseudo-Labels)

SHAP Beeswarm per Farm — Red = high value pushes toward anomaly · Blue = pushes toward normal (from Notebook 1c)

SHAP beeswarm 1c
XGBoost SHAP — Top 5 Features per Farm (Interactive)
Gradient Boosting SHAP — Agreement validates sensor importance as data property
Ensemble Alert Suppression — Before vs After 30-min Window

Ensemble Voting + Alert Suppression — Raw 2/3 ensemble (top) vs after 30-min confirmation window (bottom) (from Notebook 1c)

Alert suppression
CARE-Optimised Ensemble Results

CARE Score: Baseline vs Optimised Ensemble Threshold

XGB vs GB SHAP Agreement — Features in Top-5 of Both Models

Thermal & Electrical Signal Analysis

Notebooks 2a · 2b — Signal ranking · FFT · Correlation instability · Autoencoder

Sensor Signal Strength — Combined Score per Farm
Welch PSD — Frequency-Domain Fault Signatures

Welch PSD: Normal vs Anomaly — New spectral peaks or broadband noise increase = electrical anomaly signature (from Notebook 2a)

Welch PSD
Correlation Instability — Thermal Sensor Decoupling

Rising instability = sensors starting to decouple = early thermal warning (from Notebook 2a)

Correlation instability
Autoencoder Reconstruction — Actual vs Reconstructed

Growing gap = sensor behaving unlike anything in normal training data (from Notebook 2b)

AE overlay
Spectral & AE Summary Charts

Spectral Excess — Anomalous Frequency Content per Farm

AE Reconstruction Error — Normal vs Anomaly Mean Error

Correlation Instability — First Alert Summary
FarmEventFirst Alert StepHours Into EventMax InstabilityStatus

Practical Monitoring Strategies

Notebook 2c — Ensemble · T² attribution · Operator risk dashboard

Alert Suppression — False Positive Reduction

FP / TP Before and After 30-min Confirmation Window

T² Sensor Attribution — Which Sensors to Inspect First
Interactive Threshold Simulator

Adjust Detection Threshold — See How CARE Components Change

Coverage
Accuracy
Reliability
Earliness
CARE Score
CARE-Optimised Thresholds
FarmBaseline CAREOptimised CAREImprovementOpt. Threshold
Operator Risk Dashboard — All Anomaly Events Ranked

Severity vs Earliness — Thermal Anomaly Events

#FarmTurbineEventSeverityPersist (h)EarlinessRisk

Model Duel

Pick any two models and a farm — head-to-head CARE component breakdown

Select Two Models to Compare

Head-to-Head CARE Components Radar

Maintenance Cost Calculator

Estimate real-world savings from your model results vs naive always-alert baseline

Input Your Cost Assumptions