Semisupervised Training of Deep Generative Models for High-Dimensional Anomaly Detection.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Abnormal behaviors in industrial systems signal critical events, necessitating accurate and timely anomaly detection.
• The rarity and high cost of labeling anomalies present significant challenges to traditional detection methods.
• Deep generative models offer powerful solutions for unsupervised and semisupervised learning tasks.

Purpose of the Study:

• To introduce a novel deep generative model, Latent Enhanced regression/classification Deep Generative Model (LEDGM), for anomaly detection in multidimensional data.
• To address the limitations of existing two-stage models by proposing an end-to-end learning paradigm.
• To optimize the model for anomaly detection performance over reconstruction accuracy.

Main Methods:

• Developed LEDGM, a deep generative model that conditions label prediction on learned latent representations.
• Employed an end-to-end learning approach, differing from previous two-stage decoupled models.
• Evaluated LEDGM on synthetic and real-world datasets, including small- and large-scale data.

Main Results:

• LEDGM demonstrated improved anomaly detection performance on multidimensional data, particularly with sparse labels.
• The study confirmed the value of both labeled anomalies and normal data in semisupervised learning scenarios.
• Ablation experiments highlighted the contribution of both original input and learned latent features to LEDGM's high performance.

Conclusions:

• LEDGM offers a more effective approach to anomaly detection in industrial systems compared to prior deep generative models.
• The findings underscore the importance of leveraging available labeled data, whether normal or anomalous, for enhanced detection.
• Further labeled data generally leads to improved anomaly detection performance.