Weld pool and keyhole dynamic analysis based on visual system and neural network during laser keyhole welding

In keyhole fiber laser welding processes, the weld pool behavior and keyhole dynamics are essential to determining welding quality. To observe and control the welding process, the accurate extraction of the weld pool boundary as well as the width is required. In addition, because of the cause-and-effect relationship between the welding defects and stability of the keyhole, which is primarily determined by keyhole geometry during the welding process, the stability of keyhole needs to be considered as well. The first part of this thesis presents a weld pool edge detection technique based on an off axial green illumination laser and a coaxial image capturing system that consists of a CMOS camera and optic filters. According to the difference of image quality, a complete developed edge detection algorithm is proposed based on the local maximum gradient of grayness searching approach and linear interpolation. The extracted weld pool geometry and the width are validated by the actual welding width measurement and predictions by a numerical multi-phase model. As for the keyhole dynamics, three essential attributes to describe the simplified three-dimensional keyhole shape include keyhole size, penetration depth and keyhole inclination angle. However, when using traditional measurement techniques, it is very challenging to take in-process measurements of penetration depth and inclination angle, even if the keyhole size can be detected by using a visual monitoring system. To realize the on-line estimation of keyhole dynamics and welding defects, a data-based radial basis function neural network state observer is adopted for estimating penetration depth and inclination angle in the transient state when welding parameters change suddenly. First, a static neural network is trained in advance to establish a correlation between the welding parameters and unobservable keyhole geometry. The dynamic state observer is trained based on the transient welding conditions predicted by a numerical model and then used to estimate the time-varying keyhole geometery. Meanwhile, the coaxial monitoring system is used to observe the keyhole shape from the top side in real time, which not only provides input to the neural network but also indicates the potential welding porosities. The predicted results are validated by experimental data performed by welding with stainless steel 304 and magnesium alloy AZ31B