Deutsch

Titel: Predictive Maintenance (Regression + Klassifikation)

Ziel:
Frühzeitige Erkennung von Störungen an Maschinenlagern und Sensordaten, bevor akustische oder physische Schäden auftreten.
Kombination aus Restnutzungsdauer-Schätzung (RUL, Regression) und Frühwarnsignal (Klassifikation).

Vorgehen:

1. GPU-Setup: Nutzung von LightGBM mit GPU-Beschleunigung
2. Datenaufbereitung: Laden von Sensordaten (CSV), Zeitstempel, Feature-Lags, Condition-Indicator
3. Zielvariablen:
   • time_to_failure (Regression)
   • y_cls_early_window (binary classification, 60-Minuten-Fenster)
   • Klassengewichte gegen Class-Imbalance
4. Feature Engineering: Skalierung, One-Hot-Encoding, Ordinal-Encoding (severity)
5. Modellierung: LightGBM Regressor + Classifier, Hyperparameter-Tuning mit HalvingRandomSearchCV
6. Cross-Validation: TimeSeriesSplit zur Sicherung zeitlicher Konsistenz
7. Metriken:
   • Regression: MAE, MAPE
   • Klassifikation: PR-AUC, F1-Score
   • Kosten-sensitive Schwellenwertoptimierung (FN=5, FP=1)
8. Visualisierungen: ROC-Kurve, Precision-Recall, Konfusionsmatrix, Feature-Importance, Residuen-Analysen
9. Artefakte:
   • Modelle (rul_lgbm.joblib, earlywarn_lgbm.joblib, preprocessor.joblib)
   • Schwellenwert (earlywarn_threshold.joblib)
   • Metriken (metrics_summary_pm2.csv)
   • Grafiken (viz/)

Ergebnis:

• PR-AUC = 0.87
• Downtime-Reduktion: 10–12 %

---

English

Title: Predictive Maintenance (Regression + Classification)

Objective:
Early detection of failures in bearings and sensor data before acoustic or physical damage occurs.
Combines Remaining Useful Life estimation (RUL, regression) with early warning classification.

Workflow:

1. GPU setup: Using LightGBM with GPU acceleration
2. Data preprocessing: Load CSV sensor data, timestamps, feature lags, condition indicator
3. Target variables:
   • time_to_failure (regression)
   • y_cls_early_window (binary classification, 60-minute window)
   • Class weights to handle imbalance
4. Feature Engineering: Scaling, one-hot encoding, ordinal encoding (severity)
5. Modeling: LightGBM regressor + classifier, hyperparameter tuning with HalvingRandomSearchCV
6. Cross-validation: TimeSeriesSplit to ensure temporal consistency
7. Metrics:
   • Regression: MAE, MAPE
   • Classification: PR-AUC, F1-score
   • Cost-sensitive threshold optimization (FN=5, FP=1)
8. Visualizations: ROC curve, precision-recall, confusion matrix, feature importance, residual analysis
9. Artifacts:
   • Models (rul_lgbm.joblib, earlywarn_lgbm.joblib, preprocessor.joblib)
   • Threshold (earlywarn_threshold.joblib)
   • Metrics (metrics_summary_pm2.csv)
   • Plots (viz/)

Result:

• PR-AUC = 0.87
• Downtime reduction: 10–12 %