Online monitoring system for welding states of bottom-locking joints in high-speed trains via multi-information fusion and 3DCNN

Recognition and control of the laser welding states of aluminum profiles in China Railway High-speed trains is a challenge to current monitoring system, because the bottom-locking joints possess more complex welding states (Lack, Lock, Good, Full and Over penetration) than common butt joints. This work developed an online monitoring system based on a 3D-convolutional neural network (3DCNN), which utilized temporal information of both laser plume and weld pool images. The results indicated that multi-information fusion was necessary because the ability of single-information method to distinguish five welding states were insufficient. On this basis, there kinds of 3DCNN using different information-fusion methods were compared. The separated-channel 3DCNN with feature-level fusion was found to be more efficient than the original 3DCNN with data-level fusion. It realized the same level accuracy while decreasing the network params by 49.5 %. Additionally, the importance of different information was considered by trainable weight indexes and weight-fusion 3DCNN was proposed. The weight of weld pool information gradually decreased while the weight of laser plume increased. The weight-fusion 3DCNN was superior to the other two networks and the highest accuracy of 98.9 % was obtained. The processing speed of the whole online monitoring system with feedback control reached 10 Hz, which could meet the requirements in industrial applications. This work will provide more guidance on the application of multi-information fusion and deep learning methods in online monitoring system.