Estimation of Elbow Wall Thinning Using Ensemble-Averaged Mel-Spectrogram with ResNet-like Architecture

Area of Science:

• Mechanical Engineering
• Non-Destructive Testing
• Vibration Analysis

Background:

• Wall thinning in industrial pipe elbows is a critical issue affecting structural integrity.
• Early detection of wall thinning is essential for preventing catastrophic failures and ensuring operational safety.
• Traditional inspection methods can be time-consuming and may not detect localized thinning effectively.

Purpose of the Study:

• To propose a novel method for diagnosing elbow wall thinning.
• To utilize stationary vibration characteristics for defect detection.
• To develop a predictive model for quantifying wall thinning thickness.

Main Methods:

• Measurements of accelerations on curved pipe surfaces within a closed test loop.
• Analysis of vibration characteristics using mel-spectrograms to identify modal variations.
• Application of an ensemble mean of mel-spectrograms to enhance stationary signals and reduce noise.
• Development of a convolutional neural network (CNN) regression model incorporating residual blocks.

Main Results:

• The proposed method successfully extracted vibration characteristics indicative of wall thinning.
• Using the ensemble mean of mel-spectrograms improved model prediction accuracy by emphasizing stationary signals.
• The CNN regression model with residual blocks demonstrated superior performance in predicting wall thinning.
• The model accurately predicted the thinning thickness of elbows not included in the training dataset.

Conclusions:

• The developed method offers an effective approach for diagnosing elbow wall thinning based on vibration analysis.
• The integration of mel-spectrograms and CNN regression with residual blocks enhances diagnostic accuracy.
• This technique holds potential for real-time, non-destructive monitoring of pipe integrity in industrial applications.