Application of artificial neural network in laser welding defect diagnosis

In this paper, audible sounds during keyhole and conduction laser welding were analyzed. The characteristic signals representing good welding quality was from 10 to 20 kHz. The more the welded metal vaporizes, the higher the plasma temperature and the stronger the acoustic signals. Furthermore, keyhole shape also affected the acoustic signal intensities. Then time domain, frequency domain and wavelet analysis methods were used to analyze the acoustic signals. It was proved that frequency distributions are a better way to identify welding defects. The wavelet analysis results showed that the intensity of low frequency (<781 Hz) components of the sound signals decreased dramatically when welding defects occurred. At the end, an artificial neural network (ANN) was constructed to diagnose welding faults. Features extracted from the acoustic signals were input into the ANN. After training, the ANN could be used to identify between normal and abnormal welds.