Improved Convolutional Neural Network for Laser Welding Defect Prediction

다국어 초록 (Multilingual Abstract)

In order to predict the laser welding defects, a convolutional neural network prediction model is established. The keyhole image and plume image collected by a high-speed camera are processed to obtain visual information such as keyhole area and plume area. The rolling mean and standard deviation methods are used to calculate the fluctuation degree indicators of the visual information and the optical radiation information obtained by the photoelectric sensor. Finally, three improved one-dimensional convolutional neural network prediction models with a learning rate dynamic adjustment mechanism are established to predict welding defects. Experimental results indicate that the improved one-dimensional convolutional neural network prediction model can avoid premature convergence four times to achieve the best performance. The fluctuation degree indicators of sensor features can distinguish the welding state more easily than the sensor features. The reliability test of the new weld is carried out. The prediction accuracy of fusion detection model of sensor features and fluctuation degree indicators is 99.21%. The improved model can accurately predict laser welding defects.