Exploratory Analysis
What you'll see here
What the 40 ciphered predictors look like on their own, how they shift with failure, how they correlate with each other, and how outliers were treated.
Predictors
40
all numeric, none constant
Mean outlier share
1.25%
IQR rule, retained not clipped
Missing cells
36 / 800,000
0.005% of the matrix
Strongest signal
V18
corr -0.293 with failure
Univariate observations
- All 40 predictors are continuous and broadly bell-shaped around zero, consistent with a transformed/ciphered sensor feed.
- Several variables show heavier tails than a normal curve; on average 1.25% of values sit outside the 1.5 × IQR fence.
- No zero-variance or constant columns, so every predictor carries some signal and none can be dropped on inspection alone.
- Tails are genuine sensor extremes rather than data-entry errors, so outliers are kept — standardised inputs handle them, and extreme readings are exactly where failure signal lives.
- Only 36 missing cells appear in training (V1 and V2, 18 each) and 11 in test — a negligible share treated by median imputation.
Bivariate — correlation with failure
| Predictor | Corr. | Direction |
|---|---|---|
| V18 | -0.293 | Lower values → more failures |
| V21 | +0.256 | Higher values → more failures |
| V15 | +0.249 | Higher values → more failures |
| V7 | +0.237 | Higher values → more failures |
| V16 | +0.231 | Higher values → more failures |
| V39 | -0.227 | Lower values → more failures |
Multivariate — strongest predictor pairs
| Pair | |Correlation| |
|---|---|
| V15 – V7 | 0.87 |
| V2 – V14 | 0.85 |
| V21 – V16 | 0.84 |
| V24 – V32 | 0.83 |
| V11 – V29 | 0.81 |
- The strongest pairwise correlation is 0.87 — a handful of sensor pairs move together, most likely readings from the same subsystem.
- The rest of the matrix is close to zero, so predictors are largely independent and no aggressive dimensionality reduction is needed.
- Mild multicollinearity is not a problem for neural networks the way it is for linear models — no variables were dropped.
- Keeping all 40 inputs preserves the interaction effects the hidden layers are there to learn.
Bivariate & outlier read
- V18 is the strongest single signal (-0.29): failing generators sit systematically lower on this sensor.
- V21, V15, V7 and V16 shift upward for failures; V39, V36 and V3 shift downward.
- No single predictor separates the classes on its own — the shifts overlap heavily.
- The IQR scan flags an average of 1.25% of values per predictor; they are retained deliberately, because extreme sensor readings are informative for failure detection.
- This is exactly the case for a neural network: the failure signature lives in the combination of sensors, not in any one of them.
AI: What the EDA reveals
Click Generate to produce an evidence-based observation, insight, recommendation, and business benefit for this section.