Abstract: Taking the penetration prediction of TIG welding as the research object, this paper proposes a penetration prediction model based on capsule network, and introduces the Squeeze-Excitation module into the model to improve the accuracy and training speed of the model. The model takes the front image of the weld pool in the welding area as the input, extracts the shallow features by using the convolution layer, adaptively extracts the importance of different features by Squeeze-Excitation module, adjusts the feature map according to the importance, and then transmits it to the capsule layer. The dynamic routing algorithm is used to update the coupling coefficient between the capsule layers, and the model finally obtains the prediction results of three penetration states (incomplete penetration, penetration and over penetration). When the number of iterations is 70, the activation function is ReLU, the number of convolution layers is 2, and the number of capsule layers is 2, the model performs best in prediction accuracy and training time. In order to further verify the performance of the proposed model, it is compared with traditional capsule network, AlexNet and VGG19. The results show that the improved capsule neural network can effectively predict the penetration state of weld, which is stable and accurate on small sample data sets.