Anomaly Detection in Asset Degradation Process Using Variational Autoencoder and Explanations.

Area of Science:

• * Industrial Engineering
• * Artificial Intelligence
• * Machine Learning

Background:

• * Predictive maintenance (PdM) is crucial for Industry 4.0, relying on data-driven machine learning.
• * A significant challenge in PdM implementation is the scarcity of high-quality labeled data.
• * Unsupervised learning offers a potential solution to overcome data limitations.

Purpose of the Study:

• * To investigate the application of unsupervised learning using variational autoencoders (VAEs) for monitoring industrial asset wear.
• * To evaluate VAE performance against a baseline autoencoder using both real-world and simulated datasets.
• * To assess the utility of explainability methods in enhancing the reliability of AI-based PdM solutions.

Main Methods:

• * Implemented unsupervised learning with a variational autoencoder for anomaly detection.
• * Utilized a real-world dataset from a hot strip mill and a simulated NASA turbofan engine dataset.
• * Employed explainability methods to interpret model predictions and understand decision-making processes.

Main Results:

• * The variational autoencoder demonstrated slightly superior performance in anomaly detection compared to the base autoencoder.
• * Performance on the industrial use-case indicates the VAE is not yet production-ready, necessitating further research.
• * Explainability methods provided insights that can enhance the trustworthiness of the AI solution.

Conclusions:

• * Unsupervised learning with VAEs shows promise for predictive maintenance, particularly in data-scarce environments.
• * While VAEs offer advantages, further development is required to meet industrial production standards.
• * Explainability is key to building confidence and reliability in AI-driven industrial monitoring systems.