To address the limitations associated with the low intelligence of welding robots, a weld seam type identification and initial point guidance method based on deep neural network named WeldNet was proposed. By incorporating channel shuffling and an attention module, the size of the WeldNet model is reduced while preserving high detection accuracy. With the help of the proposed Center-Box annotation method, the optimized WeldNet network can not only automatically identify the type of welding workpieces at a frequency of 66 Hz, but also extract the initial point of the weld seam with an error of less than 1.63 pixels. Based on the principle of "monocular vision dual position shooting", automatic guidance of the initial point of the weld seam is achieved, which greatly improves the intelligence level of welding robots. The experimental results show that the method proposed can accurately identify various types of weld joints such as butt joints, lap joints, and fillet joints with a recognition rate of 99.6%, and the method can also guide the welding torch to align with the initial point of the weld seam with an error of just 0.85 mm.