Abstract: The visual image of welding pool contains rich information of the formation of welding pool and weld seam.However, due to the complex state of welding pool and numerous interference factors, it is very difficult to design the image processing algorithm and realize the engineering application. The convolutional neural（VGG） network of the deep learning technology was used to realize the penetration state detection of thin plate TIG welding. The VGG network was firstly used to build a model to judge the penetration state from the visual image of the reverse side of the molten pool, which realizes the reliable identification of non-penetration, penetration and burn-through with the accuracy of 97.2%. On this basis, SSD network was also used to build a detection model for the penetration state, which realizes the accurate measurement of the reverse side of the molten pool. In addition, the method of data enhancement was used to simulate different detection conditions so that the model’s adaptability reaches the engineering level. At the same time, a network model which predicts the penetration of the reverse side from the positive molten pool was constructed, which solves the situation that can not be directly judged from the opposite side, and the prediction accuracy of the model is 96.7%. Finally, the causes of errors and the methods to improve the accuracy were analyzed.