Abstract: Objective In order to further improve feature completeness and correlation expression of depth feature enhancement model for ultrasonic signals of welds and global optimization efficiency of model parameter adaptation, a weld defect detection method based on lightweight feature enhancement network was carried out in this paper. Methods By constructing weld defect detection model, correlation characterization of weld characteristics and different spatial features was enriched. Sparrow search algorithm based on global optimization strategy was introduced to further improve efficiency and performance of parameter adaptive optimization. At the same time, the improved optimization algorithm was used to self-optimize four key model parameters required by the constructed model, and finally the detection model suitable for spatial characteristics of weld defects with self-learning ability was constructed. Results The experimental results showed that the proposed model had a recognition accuracy of 95.54%, and average test time of a single sample was only 1.4 ms, which achieved better detection effect than other baseline models, and could meet requirements of online real-time identification of stainless steel weld defects, verifying its effectiveness and generalization. Conclusion This method reduced impact of artificial network design on parameters, performance, learning ability and cost consumption, and could be widely used in researching automatic model construction in different defect detection industries, which provided favorable technical support and guarantee for industrial modernization.