Main Problem

Lack of detecting unseen fault samples of DL models during test operations

Since collecting data for all types of faults is challenging, DL models are often trained with only a subset of fault modes
→ when unseen fault data is used for testing, it may be misclassified as another fault type or even as normal
Conventional methods that detect unseen faults (Bayesian Neural Network, Deep Ensemble) show low detection performance.

Low performance of Conventional Uncertainty-aware DL models (Bayesian Neural Network, Deep Ensemble)

Feature is not discriminative enough to detect unseen data (large intra-class variance)
→ The model needs to learn not only what is different, but also the similarity of the features
For conventional models, uncertain data should be mapped into a small region where all class probability is the same. But it doesn’t happen in practice
→ Uncertain data that can be embedded in various feature spaces must be considered

Page updated

Google Sites

Report abuse