Real-time weld geometry prediction based on multi-information using neural network optimized by PCA and GA during thin-plate laser welding

Real-time monitoring of the welding quality is quite important during the process of industrial laser manufacturing. In this paper, a multi-information fused neural network, combining welding parameters and morphological features of the molten pool, was proposed to predict geometric features of the weld seam. Firstly, a modified optical fiber laser coaxial monitoring platform was set up to acquire clear images of the molten pool. Then, several morphological characteristics of the molten pool were extracted. By using principal component analysis (PCA) to reduce the redundancy of these features, the welding speed, the laser power and the two PCA components acted on as the four input neurons, while the two output neurons consisted of the weld waist width (WW) and the weld back width (BW) representing weld seam quality. Before training, the genetic algorithm (GA) was adopted to optimize the initialized weights and bias of the neural network due to its globally search ability. The experiment results showed that our proposed model can effectively and steadily predict the geometric features of the weld seam with the mean absolute percentage error (MAPE) less than 1% and the mean square error (MSE) less than 10(-3). Time analysis showed that the whole process time of our system containing feature extraction and neural network was less than 90 ms which can meet the time requirements of large-scale real-time thin-plate laser welding application. Our system lays a foundation on the real-time quality monitoring in the process of laser welding thin-plate butt joint.