A fusion sparse learning algorithm for fault identification of rolling bearings

Area of Science:

• Mechanical Engineering
• Data Science
• Artificial Intelligence

Background:

• Rolling bearings are critical components in CNC machine tools, necessitating effective data-driven fault diagnosis.
• Existing deep learning models can be complex and lack interpretability for edge device deployment.
• A need exists for efficient and accurate fault diagnosis algorithms for rolling bearings.

Purpose of the Study:

• To propose a two-stage fusion sparse learning algorithm for rolling bearing fault diagnosis.
• To leverage Long Short-Term Memory (LSTM) for temporal feature extraction and sparse learning on Stochastic Configuration Networks (SCN) for classification.
• To develop a lightweight and interpretable model suitable for edge devices.

Main Methods:

• Feature extraction using Long Short-Term Memory (LSTM) networks to capture temporal characteristics of sensor data.
• Classification using a novel sparse learning algorithm with L0 regularization on Stochastic Configuration Networks (SCN).
• Iterative learning formula combining ADMM and quadratic equations theory, with an inequality supervision mechanism.

Main Results:

• The proposed LSTM-L0-SCN algorithm achieved optimal sparsity degrees of 76.66% on a benchmark dataset and 29.39% on the CWRU dataset.
• Demonstrated a 30% improvement in sparsity compared to the Pooling-based Sparse Coding Network (PSCN).
• Achieved an optimal test classification accuracy of 97.51% on the CWRU dataset, validating its effectiveness in rolling bearing fault identification.

Conclusions:

• The developed two-stage fusion sparse learning algorithm effectively diagnoses rolling bearing faults.
• The algorithm combines LSTM's temporal feature extraction with SCN's sparsity and efficiency, outperforming existing methods.
• The model's lightweight nature and interpretability make it suitable for deployment on edge devices for real-time monitoring.