BiVi-GAN: Bivariate Vibration GAN.

Area of Science:

• Engineering
• Artificial Intelligence
• Data Science

Background:

• Prognosis and Health Management (PHM) for rotating machinery is critical for equipment reliability.
• Current AI and deep learning methods in PHM require substantial data, posing practical challenges.
• Traditional deep learning models often struggle to capture intrinsic physical characteristics vital for vibration analysis.

Purpose of the Study:

• To develop a novel approach for augmenting vibration data in PHM using a specialized generative model.
• To address the data-intensive limitations of current AI models in PHM.
• To improve the accuracy and robustness of PHM for rotating machinery.

Main Methods:

• Introduction of the bivariate vibration generative adversarial networks (BiVi-GAN) model.
• Integration of physics-informed neural network (PINN) principles into the generative model.
• Incorporation of physical information (PI) through order analysis and cross-wavelet transform, alongside PI loss.

Main Results:

• The BiVi-GAN model demonstrated effectiveness in augmenting vibration data for PHM.
• The model successfully incorporated domain-specific physical insights into a data-driven AI framework.
• A 70% performance improvement in JS divergence was observed compared to the baseline biwavelet-GAN model.

Conclusions:

• The proposed BiVi-GAN model offers a robust and accurate solution for PHM in rotating machinery.
• Combining domain-specific knowledge with AI enhances the performance of data-driven models.
• This approach holds significant potential for improving equipment reliability through advanced PHM.