Robust Sparse Non-Negative Matrix Factorization for Identifying Signals of Interest in Bearing Fault Detection

Area of Science:

• Mechanical Engineering
• Signal Processing
• Machine Learning

Background:

• Bearings are critical components in rotating systems, and their early fault detection is essential for industrial applications.
• Classical fault detection methods struggle with heavy-tailed or non-cyclic impulsive noise.
• Deep learning methods often require large labeled datasets and assume Gaussian noise, which are not always practical.

Purpose of the Study:

• To develop a robust fault detection method for rotating machinery that can handle heavy-tailed noise.
• To improve the accuracy of fault frequency band identification in bearing signals.
• To address the limitations of existing classical and deep learning approaches in noisy industrial environments.

Main Methods:

• A sparse non-negative matrix factorization (NMF) method was developed using the maximum-correntropy criterion for robustness against heavy-tailed noise.
• The proposed NMF method was applied to identify fault frequency bands within the signal's spectrogram.
• Monte Carlo simulations and statistical efficiency analysis were used to validate the method's performance under various noise conditions.

Main Results:

• The proposed method demonstrated effectiveness in identifying fault frequency bands in simulated signals with both Gaussian and heavy-tailed noise.
• Statistical analysis confirmed the method's robustness against random perturbations.
• Evaluation on three real-world datasets showed the practical applicability and effectiveness of the approach.

Conclusions:

• The robust sparse NMF method based on the maximum-correntropy criterion offers a promising solution for early fault detection in rotating systems, particularly in the presence of challenging noise conditions.
• The method provides a reliable alternative to traditional techniques and deep learning approaches that are sensitive to noise and data availability.
• The findings highlight the potential for improved machinery diagnostics and predictive maintenance in industrial settings.