Feature extraction in welding penetration monitoring with arc sound signals

Abstract

Using welding sound signals to monitor the welding pool state has potential in gas tungsten arc welding. A novel welding-penetration-monitoring method using the welding sound based on feature extraction and feature selection is proposed. In order to overcome the blindness of subjective dimensionless indicators selected as sensitive features without any experience, attempts are made to try to obtain as many relative features as possible. A feature selection technology is used to find the most effective features and to reduce the redundant features. In this paper, a wavelet packet transform method is used to decompose the welding sound signals; then the decomposed nodes are calculated and 128 features obtained through statistical processing; a genetic algorithm selects seven feature subsets from the 128 features and employs an artificial neural network to classify the different penetration states, with 85 per cent accuracy in test data.