An Intelligent Ball Bearing Fault Diagnosis System Using Enhanced Rotational Characteristics on Spectrogram.

Area of Science:

• Mechanical Engineering
• Signal Processing
• Artificial Intelligence

Background:

• Ball bearing faults are a primary cause of rotating machinery failure.
• Existing fault diagnosis methods like temperature and sound monitoring have limitations due to external noise.
• Vibration monitoring, particularly using mel-frequency cepstral coefficients (MFCCs) and 2D convolutional neural networks (CNNs), is a promising alternative but faces challenges with high sampling rates, broad frequency utilization, and model complexity.

Purpose of the Study:

• To address the limitations of existing vibration-based bearing fault diagnosis methods.
• To propose a novel fault diagnosis system that is accurate, computationally efficient, and adaptable to changing operational conditions.
• To reduce the complexity and resource requirements of bearing fault detection systems.

Main Methods:

• Development of a rotational characteristic emphasis (RCE) spectrogram process for analyzing narrow frequency bands and generating low-resolution images.
• Design of an optimized, shallow-structure convolutional neural network (CNN) for efficient fault classification.
• Validation of the proposed system using experimental data.

Main Results:

• The proposed system achieved a high diagnostic accuracy of 0.9974.
• The optimized CNN model demonstrated low complexity with only 5.81 KB of parameters and 1.53×10^6 FLOPs.
• The system effectively handles low sampling rates and varying rotation frequencies.

Conclusions:

• The proposed RCE spectrogram and optimized CNN offer a highly accurate and low-complexity solution for ball bearing fault diagnosis.
• This approach overcomes the limitations of traditional MFCC and complex 2D CNN methods.
• The developed system is practical for real-world applications requiring adaptable and efficient machinery health monitoring.