Acoustic deep learning for defect detection in aluminium wheel rims

Area of Science:

• Materials Science
• Artificial Intelligence
• Acoustics

Background:

• Visual inspection of aluminum wheel rims for defects is unreliable and operator-dependent.
• Current diagnostic methods lack automation and precision for detecting subtle structural issues.

Purpose of the Study:

• To develop an automated diagnostic framework, RimNet, for assessing the condition of aluminum wheel rims.
• To utilize acoustic responses and deep learning for defect detection.
• To enhance the reliability and efficiency of wheel rim diagnostics.

Main Methods:

• Developed RimNet, a deep learning framework combining Short-Time Fourier Transform (STFT) with Convolutional Neural Networks (CNNs).
• Incorporated domain filtering and hard-negative mining to improve defect classification.
• Utilized a dataset of 930 labeled acoustic responses from wheel rims with various defects and serviceable conditions.

Main Results:

• RimNet achieved 96.9% accuracy at the file level in classifying wheel rim conditions.
• No wheel-level misclassifications were observed under the study's evaluation protocol.
• Analysis of uncertainty measures correlated prediction stability with defect types, identifying geometric variability as a challenge.

Conclusions:

• The proposed acoustic deep learning approach offers a reliable method for automated wheel rim defect detection.
• RimNet can be easily integrated into existing workshop environments due to its non-invasive, microphone-based data acquisition.
• This technology shows significant potential for structural health monitoring of lightweight mechanical components.