Mel frequency cepstral coefficient temporal feature integration for classifying squeak and rattle noise

Area of Science:

• Acoustics and Signal Processing
• Mechanical Engineering
• Machine Learning for Diagnostics

Background:

• Mechanical fault diagnosis increasingly relies on analyzing emitted noise.
• Efficient classification of acoustic features is crucial for accurate fault identification.
• Mel frequency cepstral coefficients (MFCCs) are effective for capturing sound characteristics.

Purpose of the Study:

• To introduce a method for selecting statistical indicators to integrate MFCC feature sets.
• To identify the most effective feature sets for classifying squeak and rattle sounds.
• To lay the groundwork for a vehicle faulty sound recognition algorithm.

Main Methods:

• Creation of specialized audio datasets for squeak and rattle sounds.
• Experimental data collection comprising 256 rattle and 144 squeak recordings across multiple classes.
• Evaluation of classifier accuracy using individual feature sets with a support vector machine (SVM).

Main Results:

• Identification of the best-performing statistical feature sets for squeak and rattle audio classification.
• Demonstration of the efficacy of MFCCs and selected statistical indicators in sound analysis.
• Validation of the support vector machine's performance in distinguishing fault sounds.

Conclusions:

• The developed method effectively selects statistical indicators for integrating MFCC features.
• The findings provide crucial insights into optimizing acoustic feature sets for fault diagnosis.
• This research serves as a foundation for advanced vehicle noise-based fault recognition systems.