Abstract: The root hump defect is easy to appear in the laser-MIG composite welding process. In order to realize the simultaneous prediction of the root hump defect in the welding process, this paper studies the algorithm of root hump defect prediction and analyzes the prediction results of different algorithm. The real-time visual sensing information of composite welding process is carried out by a high-speed camera, the time series characteristic information of the front weld pool and the keyhole in the welding process is extracted, and the characteristics signals are decomposed and reconstructed by wavelet packet decomposition (WPD). Then, the residual height of the back weld is obtained by a laser scanner, which is used as the basis for marking the hump status. Long short-term memory (LSTM) neural network was used to predict the status of root hump in the welding process. Experimental results show that the accuracy of WPD-LSTM algorithm for root hump prediction is 97.85%. Compared with other algorithms, the prediction accuracy of WPD-LSTM algorithm based on the temporal feature information of the front visual sensing in the welding process is higher, and the prediction results have higher continuity, which is conducive to realize the synchronous detection and control of root hump defects in welding process.