Semi-supervised deep architecture for classification in streaming data
with emerging new classes: application in condition monitoring
Abstract
Classification in streaming data with emerging new classes (SENC) is
challenging because the surrogate models trained on the closed datasets
are open to new information never seen before. The fundamental sub-tasks
include: (1) Pattern recognition of these already known classes with
high accuracy; (2) Timely identification of emerging new classes; (3)
Model update to adapt to new classes and (4) distinguish different new
classes. Existing SENC frameworks utilize predefined threshold to set
the ambiguous decision boundaries between the known and emerging new
classes, they have poor performance in distinguishing different emerging
new classes. To this end, a novel SENC framework leveraging Generative
Adversarial Network (GAN) and history-state ensemble method is proposed
in this paper, which is referred to as ensemble multi-classification GAN
(EMC-GAN). Compared with traditional SENC frameworks, the proposed
method is designed based on deep learning, thus, it can automatically
explore implicit features from the raw data without human intervention.
Besides, it doed not require pre-set threshold for novelty detection.
The proposed method is evaluated on the condition monitoring of rotating
machinery - a typical SENC case in real-world industry. Experimental
results show that the proposed SENC framework can not only effectively
detect emerging new classes from known ones, but also exhibit excellent
performance in distinguishing different new classes.