Health indicator construction by quadratic function-based deep convolutional auto-encoder and its application into...

Area of Science:

• Mechanical Engineering
• Data Science
• Artificial Intelligence

Background:

• Bearing failures can cause catastrophic machinery failures, necessitating accurate Remaining Useful Life (RUL) prediction.
• Effective Health Indicator (HI) construction is vital for data-driven RUL prediction, but current methods often require prior knowledge or struggle with raw vibration signals.

Purpose of the Study:

• To develop a novel method for constructing Health Indicators (HIs) from raw bearing vibration signals without prior knowledge.
• To enhance the accuracy and predictive power of Remaining Useful Life (RUL) predictions for bearings.

Main Methods:

• A quadratic function-based Deep Convolutional Auto-Encoder (DCAE) neural network was developed.
• Raw bearing vibration signals were preprocessed using low-pass filtering before inputting into the DCAE.
• The DCAE was used to construct Health Indicators (HIs) from the cleaned vibration data.

Main Results:

• The proposed quadratic function-based DCAE method demonstrated superior performance in constructing HIs compared to traditional methods like AE, DNN, KPCA, ISOMAP, PCA, and VAE.
• The constructed HI exhibited stronger predictive power for RUL than existing data-driven HIs.
• Experimental results validated the effectiveness of the novel HI construction approach.

Conclusions:

• The quadratic function-based DCAE offers an effective approach for Health Indicator construction from raw bearing vibration signals.
• This method overcomes limitations of existing techniques by not requiring prior knowledge and directly processing raw data.
• The enhanced HIs significantly improve the accuracy of Remaining Useful Life (RUL) predictions in machinery.