Abstract: To improve the accuracy of predicting the molten pool penetration state in plasma arc welding so as to meet industrial needs, this paper proposes a model called PCSCNet that integrates image space and channel characteristics. In this model, the convolutional residual network ResNet50 structure is modified and integrated into the channel attention network squeeze and excitation network to simultaneously extract spatial feature information and channel feature information from the front image of the molten pool. By testing on a dataset of constant current plasma arc welding experiments, the model establishes the corresponding relationship between the front surface image of the weld pool and the state of the keyhole. The results show that the model achieves a prediction accuracy of over 95%. Using the Grad-CAM method, the model's predicted focus area is visualized, analyzed, and compared with the actual molten pool's image features to verify the model's reliability.