From Envelope Spectra to Bearing Remaining Useful Life: An Intelligent Vibration-Based Prediction Model with...

Area of Science:

• Mechanical Engineering
• Machine Learning
• Reliability Engineering

Background:

• Bearings are critical in rotating machinery, and their failures can cause systemic issues.
• Accurate remaining useful life (RUL) prediction is vital for effective predictive maintenance strategies.
• Existing data-driven RUL methods lack physical interpretability and robust uncertainty quantification.

Purpose of the Study:

• To develop a robust and interpretable method for predicting bearing RUL.
• To quantify uncertainty in RUL predictions for improved maintenance planning.
• To leverage physical insights from bearing degradation for enhanced RUL estimation.

Main Methods:

• Utilized a convolutional variational autoencoder for regression (CVAER) model.
• Employed enhanced average envelope spectra (AES) as input for improved fault detection and physical robustness.
• Formulated a probabilistic regressor and latent generator for uncertainty quantification and meaningful latent space learning.

Main Results:

• The CVAER model probabilistically predicts RUL distributions with associated confidence measures.
• The use of AES isolates bearing-specific information, enhancing RUL prediction accuracy.
• Experimental validation confirmed the model's effectiveness compared to benchmark methods.

Conclusions:

• The proposed CVAER approach offers a physically interpretable and robust method for bearing RUL prediction.
• Integrating envelope spectra with deep learning enhances the reliability of predictive maintenance.
• This work advances the state-of-the-art in bearing condition monitoring and RUL estimation.