CycleStyleGAN-Based Knowledge Transfer for a Machining Digital Twin.

Area of Science:

• Manufacturing Engineering
• Data Science
• Artificial Intelligence

Background:

• Industry 4.0 relies on digitalization and digital twins for process optimization.
• Implementing digital twins faces challenges with heterogeneous and dynamic cyber-physical systems.
• Existing research offers frameworks but lacks focus on practical implementation for data analysis.

Purpose of the Study:

• To address challenges in analyzing data from heterogeneous and dynamic Industry 4.0 systems.
• To propose a digital twin simulation tool for capturing and adapting machining vibration signals.
• To enable knowledge extraction and reuse from existing manufacturing simulation models.

Main Methods:

• Development of a domain adaptation algorithm based on generative adversarial networks.
• Implementation of a novel CycleStyleGAN architecture extending CycleGAN with style-based signal encoding.
• Validation through an experimental scenario replicating a real-world manufacturing knowledge transfer problem.

Main Results:

• The proposed digital twin simulation tool successfully captures and adapts machining vibration signal dynamics.
• The CycleStyleGAN model facilitates knowledge transfer between different manufacturing environments.
• A significant reduction in the required target domain data (by one order of magnitude) was achieved.

Conclusions:

• The proposed approach offers a flexible method for knowledge extraction from diverse manufacturing simulation models.
• This enables the reuse of costly established systems, providing economic benefits to manufacturing businesses.
• The study demonstrates a viable process optimization framework for knowledge transfer in Industry 4.0.