Generative Pseudorehearsal Strategy for Fault Classification Under an Incremental Learning
Subin Lee, Jun‐Geol Baek · 2019
As fault classification becomes more important in manufacturing industry, the state-of-art machine learning methods have been utilized. However, owing to the problem called catastrophic forgetting, the networks tend to forget the former knowledge. Thus, it is evident that overall classification performance has fallen, when training the existing model with new classes. We propose classification model that retains previous information using generative pseudorehearsal networks. In this method, newly arrived fault classes could be trained on same network which is parameterized by former data. The proposed method shows significant experimental results comparing to non-incremental methods, while achieving memory efficiency and solving the class imbalance problem.