Abstract: The accurate identification of the intersection between girth and spiral welds serves as a foundation for precisely locating target pipeline segments, highlighting the necessity for an efficient and intelligent identification method. To address the challenge of intelligent identification and localization of weld intersections in long-distance pipelines, a method based on an improved Residual Neural Network 18(ResNet18)was proposed. Starting from the characteristics of magnetic flux leak-age signals at weld intersections, this method integrated an Efficient Channel Attention(ECA)mechanism into ResNet18 to create a ResNet18-ECA. It analyzed the distribution of flux leakage fields at weld intersections, combining channel information and sensor number from manual interpretations, as well as the varying effects of axial, circumferential, and radial components of pipeline magnetic flux leakage data. Additionally, the Huber loss function and Leaky ReLU activation function replaced the original MSE loss function and ReLU activation function to enhance the localization of weld intersections. Weld intersection data sets were selected for 350 training rounds, and the trends in training loss and root mean squared error (RMSE) were compared. The test results demonstrated that the improved ResNet18 performed effectively in locating weld intersections. In 20 untrained magnetic flux leakage data samples, the prediction error for the sampling point position, as determined by the weld intersection sensor number and sampling point coordinate prediction model, was within ±20 mm, while the prediction error for the sensor number was within ±84 mm. This model effectively identifies weld intersections, meets engineering requirements, and accelerates the identification process, demonstrating strong potential for application in in-line inspection data processing.