Abstract: Objective In order to solve the problem that the weld images collected in the real-time welding process are easily disturbed, a new weld seam tracking method is proposed to improve the tracking accuracy and robustness for the multi-layer and multi-pass weld seam tracking technology of process pipelines. Methods In this paper, a weld seam tracking method based on deep regression network was proposed. This method combines visual sensor calibration technology to realize weld seam tracking and control. The core is to use the features of each layer of the convolutional neural network to determine the feature points of the weld through feature similarity matching, and use the weld detection network to periodically initialize the target filter. Results The experimental results show that in the face of strong interference, the weld seam tracking accuracy of the proposed algorithm can reach 0.4 mm, which is better than other comparison algorithms. This method is especially suitable for multi-layer and multi-pass weld seam tracing of butt pipes. Conclusion The proposed method based on deep regression network effectively solves the interference problem in real-time weld image acquisition. By improving the tracking accuracy and robustness, especially the high accuracy of 0.4 mm under strong interference conditions, it provides an effective solution for multi-layer and multi-pass weld seam tracking of process pipelines.