Intelligent Fault Diagnosis of Industrial Bearings Using Transfer Learning and CNNs Pre-Trained for Audio...

Area of Science:

• Engineering
• Artificial Intelligence
• Machine Learning

Background:

• Training Artificial Intelligence (AI) algorithms for machine diagnosis typically requires extensive datasets, which are often scarce in industrial settings.
• Convolutional neural networks pre-trained for audio classification possess transferable knowledge applicable to analyzing bearing vibration spectrograms.
• Both audio classification and bearing vibration analysis necessitate the extraction of relevant features from spectrograms.

Purpose of the Study:

• To investigate the efficacy of transfer learning using pre-trained audio classification models for diagnosing localized defects in rolling element bearings.
• To demonstrate a method for knowledge transfer that minimizes the data requirements for fine-tuning AI models in diagnostic applications.
• To adapt the VGGish model for classifying bearing conditions using vibration data.

Main Methods:

• Utilized transfer learning by fine-tuning the VGGish model, pre-trained on audio classification tasks, for analyzing bearing vibration spectrograms.
• Collected vibration data from a dedicated test bench featuring medium-size bearings with induced localized defects across three damage classes.
• Employed spectrogram analysis as the common feature extraction technique for both audio and vibration data.

Main Results:

• The VGGish model, pre-trained on sound spectrograms, successfully classified bearing conditions using vibration spectrograms.
• The transfer learning approach demonstrated significant effectiveness in identifying localized defects in rolling element bearings.
• The fine-tuned model achieved accurate classification, highlighting the shared feature extraction needs between audio and vibration analysis.

Conclusions:

• Convolutional neural networks pre-trained for audio classification contain valuable knowledge transferable to bearing vibration diagnosis.
• Transfer learning offers a powerful solution to overcome data scarcity in industrial machine diagnosis, enabling efficient model fine-tuning.
• This study validates the successful application of audio-trained AI models for effective bearing health monitoring through vibration analysis.