Apps · Features
Feature Space Explorer
A classifier can only separate what the features separate. Choose features, project them to two dimensions and see which fault classes stand apart and which ones blur together.
Feature space
How well each feature separates the classesFisher score
Correlation between selected featuresPearson r
Strongly correlated features (dark cells) carry overlapping information. Keeping both adds little.
Reading the views
- PCA rotates the standardized features so the first two axes keep as much spread as possible. Classes that sit apart here are easy for almost any classifier.
- Two features shows raw values. Try kurtosis against envelope at BPFO: the outer-race defects shoot away from the rest. Try 2×/1× against 1× amplitude for misalignment versus unbalance.
- The Fisher score compares the spread between class means with the spread inside each class. High means the feature alone separates the classes well.
- The kNN check classifies each window by its five nearest neighbours among the others, using only the selected features. It is a quick estimate of how much information your feature set carries.
Windows from the same machine are similar, so the kNN check here is optimistic. The Fault Classifier shows an honest split by machine. Read more in Data and feature engineering.