Short-Time Fourier Transform and Neural Network Analysis for Welding Defect Classification Based on Current Signal Features

Welding quality plays a critical role in ensuring structural integrity and safety across manufacturing industries. This study introduces a signal-based approach for classifying welding defects by analyzing welding current signals using time-frequency analysis. The Short-Time Fourier Transform (STFT) was applied to extract five key frequencydomain features: peak frequency, spectral entropy, frequency centroid, mean power, and standard power spectrogram. These features were statistically validated using ANOVA and Tukey post-hoc analysis, confirming significant differences between good welds and multiple defect types. An Artificial Neural Network (ANN) classifier was developed and tested across five hidden layer configurations. The highest classification accuracy of 72.82% was achieved using five hidden neurons. However, the model demonstrated low performance in minority defect classes, with a precision of 7.30%, recall of 10.00%, and F1-score of 8.44%. These findings underscore the potential of combining STFT-based features with ANN for welding defect classification, while also highlighting limitations related to class imbalance and feature representation.