Aimed at the problem that the seam tracking system with low adaptability is sensitive to noise in the actual welding environment, and combined with the strong feature expression ability and self-learning function of the depth convolution neural network, a welding seam detection and tracking system based on depth hierarchical feature is studied. The location of seam from noise-contaminated serial images is accurately determined by this system. A fuzzy immune self-adaptive intelligent tracking control algorithm is designed to completely solve the chattering problem of welding torch following the calculated trajectory. The experimental results show that, under the interference of strong arc and splash, the metrical frequency of sensor can be up to 20 Hz, the tracking accuracy of the welding seam is about 0.2060 mm, and the end of the welding torch runs smoothly during the process of welding. The system can realize real-time tracking of the welding seam, has strong anti-interference ability, and can accurately track the trajectory of the welding seam, which can meet the requirements of welding application.