Professor Review
A rubric-aligned executive summary of the wind turbine generator failure-prediction project, with evidence links to every page that produced each result.
Headline numbers
Validation recall and F1 across all ten configurations
M8 — 128-64-32, dropout 0.3, Adam, class weights — leads on recall at 91.89% while holding precision above 91%.
Rubric coverage
Univariate grid for all 40 predictors, bivariate splits, correlation heatmap, outlier scan.
Duplicate/missing checks, median imputation, scaling and stratified split — training data only.
Recall justified from the replacement ≫ repair > inspection cost ordering.
Single hidden layer with SGD; 58.56% validation recall diagnosed as underfitting.
Ten configurations across depth, optimizer, dropout, class weights and oversampling.
M8 selected; 87.59% test recall, ROC-AUC 0.944, confusion matrix on 5,000 unseen units.
Deployment playbook, threshold policy, and limitations with future work.
Every criterion mapped to a page and a deck slide.
Reproducibility
- Train / validation split: 80 / 20, stratified on Target, random_state 42
- Imputation: median, fitted on the training split only
- Scaling: StandardScaler, fitted on the training split only
- Hidden activation / init: ReLU with He-normal initialisation
- Output / loss: sigmoid with binary cross-entropy
- Epochs / batch size: 30 / 128 for every configuration
- Class weights: failure 9.009, no failure 0.529
- Decision threshold: 0.5 for all reported metrics
- Random seed: 42 (numpy, TensorFlow and Keras)