Case File
Log:banbury-fault-detection·
Banbury Fault Detection & Model Comparison
Trained several classifiers on the same labeled set; selection used metrics and ROC behavior.
- Domain
- Predictive Maintenance
- Focus
- Model comparison, fault detection
- Stack
- Python, Scikit-learn, Pandas
- Status
- Completed
- Year
- 2025
Case cover
Confusion matrix — selected model
Problem
Sensor series are noisy; fault class is rare. A poor model both alarms falsely and misses events.
Approach
Time-window features (mean, std, slope). Stratified split; precision/recall/F1 and ROC-AUC.
Technical Structure
Data preparation
Raw sensor streams windowed; labels aligned with production logs.
Feature logic
Per-window statistical summaries and trend; kept feature count low to limit overfitting.
Model comparison
Logistic Regression, SVM, Random Forest, XGBoost; same CV strategy.
Evaluation metrics
Confusion matrix, F1, ROC-AUC; minority class weighting considered.
Outcome
Reported the model with the best precision–recall trade-off for this split; confusion matrix attached.
Supplementary figures
