Vibration-Based Wear Condition Estimation of Journal Bearings Using Convolutional Autoencoders

Area of Science:

• Mechanical Engineering
• Data Science
• Condition Monitoring

Background:

• Predictive maintenance is crucial for mechanical systems, especially journal bearings, to prevent failures and optimize operations.
• Traditional methods often require extensive labeled data or direct access to components, posing challenges for real-time monitoring.
• Vibration analysis is a key non-invasive technique for assessing equipment health.

Purpose of the Study:

• To develop a data-driven methodology for indirectly assessing journal bearing wear using vibration data.
• To explore the impact of sensor configurations, downsampling methods, and sampling rates on predictive accuracy.
• To leverage deep learning, specifically Convolutional Autoencoders (CAEs), for wear state estimation.

Main Methods:

• A novel experimental setup was used to accelerate wear on journal bearings.
• Vibration data from 17 bearings were collected using various sensors and mounting configurations.
• Feature engineering involved exploring downsampling techniques and sampling rates.
• Convolutional Autoencoders (CAEs) were employed to extract latent state vectors from processed vibration data.

Main Results:

• The extracted latent state vectors from CAEs showed a strong correlation with the actual wear state of the journal bearings.
• The CAE model achieved an average Pearson coefficient of 91% in estimating wear across four experimental setups.
• The methodology demonstrated effective wear estimation even with limited labeled training data.

Conclusions:

• The proposed data-driven methodology accurately estimates journal bearing wear using vibration data and CAEs.
• This approach offers a cost-effective and efficient solution for predictive maintenance of journal bearings.
• The findings highlight the potential of unsupervised deep learning for condition monitoring in mechanical systems.