Preprocessing
Every transformation applied to the data, and the guarantee that none of them saw the test set.
0 duplicate rows in train, 0 in test — nothing removed
36 cells in train (V1, V2) and 11 in test, filled with the training median
None applied — extreme sensor values are retained as failure signal
Not applicable: predictors are already ciphered, transformed sensor signals
Not required — no categorical variables in the dataset
StandardScaler fitted on the training split and applied to validation and test
- The 20,000 training rows are split 80/20 with
stratify=Targetso both sides keep the same 5.55% failure rate. - Class weights of 9.009 for failures and 0.529 for healthy units are computed from the training split and used in the weighted models.
- Random oversampling of the minority class is tested as an alternative to class weights, applied to the training split only.
- Validation and test keep their natural failure rate — no resampling is applied to either.
| Event | Assumed unit cost |
|---|---|
| Inspection (false positive) | $5,000 |
| Repair (true positive) | $15,000 |
| Replacement (false negative) | $60,000 |
Absolute costs are illustrative and preserve the stated ordering (replacement >> repair > inspection); the percentage saving is what should be read, not the dollar amounts.
Click Generate to produce an evidence-based observation, insight, recommendation, and business benefit for this section.