Condition Monitoring Using a Latent Space of Variational Autoencoder Trained Only on a Healthy Machine

Area of Science:

• Machine Learning
• Artificial Intelligence
• Diagnostics Engineering

Background:

• Traditional machine diagnostics often rely on discriminative models, requiring labeled data for fault identification.
• The scarcity or difficulty in obtaining labeled data for defective states poses a significant challenge in machine diagnostics.
• Deep neural networks offer advanced feature extraction and classification capabilities for improved decision-making.

Purpose of the Study:

• To propose an unsupervised approach for fault diagnosis in machines and mechanical structures.
• To leverage generative models, specifically Variational Autoencoders (VAEs), for automated machine diagnostics.
• To address the challenge of limited or unavailable labeled data in defected states.

Main Methods:

• Utilizing Variational Autoencoder (VAE) algorithms for generative modeling and feature extraction.
• Employing variational inference to approximate probability distributions of diagnostic features as random variables.
• Implementing unsupervised learning with automatic feature extraction from the VAE encoder's latent subspace, followed by dimensionality reduction using PCA or t-SNE.

Main Results:

• The proposed unsupervised approach successfully performs fault diagnosis using only healthy data.
• Automatic feature extraction from the probabilistic latent subspace of VAEs proved effective.
• The method demonstrated viability on simulation data, validating its potential for real-world applications.

Conclusions:

• Generative models, particularly VAEs, offer a powerful alternative to traditional discriminative models in machine diagnostics.
• The unsupervised approach effectively overcomes the limitations posed by the lack of labeled defect data.
• This method provides a robust solution for anomaly detection, failure prediction, and process optimization in machine health monitoring.