A Synthetic Time-Series Generation Using a Variational Recurrent Autoencoder with an Attention Mechanism in an...

Area of Science:

• Data Science
• Artificial Intelligence
• Cyber-Physical Systems

Background:

• Data scarcity is a major challenge for developing robust AI and data science models.
• Industrial Control Systems (ICS) data is often unavailable due to security and privacy concerns, limiting model development.
• ICS datasets possess complex time-series characteristics with both short- and long-term temporal dependencies.

Purpose of the Study:

• To propose a novel method for generating synthetic time-series data for Industrial Control Systems (ICS).
• To address the challenge of data scarcity in ICS environments by leveraging synthetic data generation.
• To effectively capture and model the temporal dependencies inherent in ICS data.

Main Methods:

• Developed an attention-based variational recurrent autoencoder (AVRAE) model.
• Extended the evidence lower bound for variational inference to time-series data.
• Integrated an attention mechanism within a recurrent neural network-based autoencoder to learn temporal dependencies.

Main Results:

• The proposed AVRAE model successfully generated visually and statistically plausible synthetic ICS time-series data.
• Comprehensive evaluation on the HAI ICS dataset demonstrated the effectiveness of AVRAE.
• The attention mechanism enabled effective learning of both short- and long-term temporal dependencies.

Conclusions:

• AVRAE provides a viable solution for generating high-quality synthetic ICS data, mitigating scarcity issues.
• The method enhances the potential for developing more accurate AI and data science models for ICS.
• This work contributes a novel approach to time-series data generation in critical infrastructure domains.