Abstract: Targeting the lap joint of multilayer aluminum tabs and an aluminum sheet, a real-time monitoring system for the laser welding process based on multi-band photoelectric coaxial sensing was established. Experiments on laser welding processes with different laser powers and defocusing conditions were conducted, and multi-band photoelectric signals under different laser energies were collected in real-time. Secondly, a wavelet scattering network (WSN) was used to extract multi-scale high-dimensional features from the raw signals. Combined with a long short-term memory (LSTM) network for temporal dynamic modeling, this approach ultimately achieves the goal of real-time detection of pseudo welding defects. The results indicate that, with a small sample size, the constructed WSN-LSTM model achieves an accuracy of 99.6%, and its classification performance surpasses that of other recurrent neural networks and lightweight convolutional neural network models. Additionally, the lightweight characteristic of the WSN-LSTM model results in the shortest training time, with an average processing time per sample of only 0.15 ms, making it advantageous for rapid deployment on power battery production lines and real-time detection of pseudo welding defects.