Girth weld is an important reference for the pipeline test data analysis which can be used to correct cumulative errors of odometer wheel. Thus, girth weld localization is the necessary part in test data analysis task. In this article, a lightweight convolution neural network with multi-scale receptive field is proposed, which is based on the magnetic flux leakage (MFL) data feature and the girth weld spatial distribution characteristics on the wall of pipeline. The features of girth weld can be extracted efficiently due to the axial 1D convolution with single sensor receptive field and circular ring convolution with circular global receptive field. Inspired by the label smoothing, the sample label is augmented. Moreover, some optimization design for loss function and activation function are also achieved. And a girth weld intelligent localization model is established. Finally, the model is trained and evaluated by the dataset including 5 676 samples that collected from various pipelines on-line MFL inspection data. The experiment results show that model has good convergency stability, and the test precision and recall rate reach 93. 90% and 94. 79% , respectively. Furthermore, the model is tested by MFL data of Φ610 pipeline which never participate in model training. The F1 score reaches 0. 93 which shows that the model has a good robustness and generalization ability, and has certain application engineering value.