Time-Frequency and Spectral Analysis of Welding Arc Sound for Automated SMAW Quality Classification.

Area of Science:

• Materials Science and Engineering
• Acoustics
• Artificial Intelligence

Background:

• Shielded Metal Arc Welding (SMAW) is a common but challenging process for ensuring weld quality.
• Objective assessment of weld quality in SMAW often relies on subjective visual inspection or destructive testing.
• Acoustic signal analysis presents a potential non-destructive method for real-time weld quality assessment.

Purpose of the Study:

• To investigate the feasibility of using acoustic signal analysis for classifying SMAW weld quality.
• To compare the effectiveness of time-domain acoustic signals versus time-frequency spectrograms for weld quality assessment.
• To evaluate the performance of various machine learning models in classifying welds based on acoustic data.

Main Methods:

• Collected acoustic signals during the SMAW process.
• Extracted scalar acoustic descriptors like fundamental frequency (F0) and harmonics-to-noise ratio (HNR).
• Utilized time-domain signals and time-frequency spectrograms as inputs for ten supervised machine learning models.

Main Results:

• Statistical tests confirmed significant acoustic differences between accepted and rejected welds.
• Spectrogram-based representations achieved higher accuracy (0.95-0.96) and ROC-AUC (>0.95) compared to time-domain signals.
• Machine learning models using spectrograms demonstrated low false positive (<6%) and false negative (<6%) rates.

Conclusions:

• Acoustic signal analysis, particularly using time-frequency spectrograms and machine learning, is a viable method for automated non-destructive evaluation of SMAW quality.
• Spectrograms provide a more robust representation for machine learning classification than simple scalar acoustic descriptors.
• This approach offers a reliable framework for real-time quality control in manual SMAW operations.