Diff-MTS: Temporal-Augmented Conditional Diffusion-Based AIGC for Industrial Time Series Toward the Large Model Era.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Data Science

Background:

• Industrial multivariate time series (MTS) data is vital for monitoring machine states.
• Insufficient data availability due to collection challenges and privacy concerns hinders industrial AI development.
• Generative Adversarial Networks (GANs) are commonly used for MTS generation but suffer from unstable training.

Purpose of the Study:

• To propose a novel diffusion model, Diff-MTS, for generating high-quality industrial MTS data.
• To address the limitations of existing GAN-based methods in MTS generation.
• To improve the diversity, fidelity, and utility of synthetic industrial time series data.

Main Methods:

• Developed a temporal-augmented conditional adaptive diffusion model (Diff-MTS).
• Introduced a conditional adaptive maximum-mean discrepancy (Ada-MMD) for controlled MTS generation without a classifier.
• Integrated a temporal decomposition reconstruction UNet (TDR-UNet) to capture complex temporal patterns.

Main Results:

• Diff-MTS demonstrated superior performance over GAN-based methods on C-MAPSS and FEMTO datasets.
• The proposed model achieved significant improvements in diversity, fidelity, and utility of generated MTS.
• Ada-MMD enhanced condition consistency in the diffusion model.

Conclusions:

• Diff-MTS effectively generates high-quality industrial multivariate time series data.
• The proposed method facilitates the development of industrial intelligence and large models.
• Diff-MTS contributes to advancements in intelligent maintenance and data generation for industrial applications.