Development of Anomaly Detection Model for Welding Classification Using Arc Sound

Abstract

This study introduces the method to classify weld bead type from arc sound of the gas metal arc welding process by applying machine learning techniques. In this research, we mainly focused on two types of weld bead which were normal weld bead and burn-through weld bead. The signal processing technique was implemented in this work to visualize welding sound data, recorded with a microphone array. All recorded sounds are imported for generating the spectrogram using Python programming and Fourier transformation to analyze and explore the difference of each sound that occurred from different weld bead types. The feature extraction from the sound data is used to construct the dataset for developing the model. Three machine learning models were trained by three different algorithms. Which were recurrent neural network (RNN), Long-short Term Memory (LSTM), and one-class Support Vector Machine (one-class SVM). Each model was evaluated with accuracy and confusion matrix. After a train and testing each model, the result showed that each model performs with an overall accuracy greater than 80 percent for each model. Given the performance of the model developed in this research, these models can be applied to the welding process. And the method from this research can also be applied with another manufacturing process in future work.