Objective Strike-slip faults are common geological hazards affecting buried pipelines. The action of such faults frequently induces significant local deformation and pipeline failure, posing serious risks to the structural integrity of pipeline networks. Therefore, rapid and accurate prediction of strain peaks for pipelines under the action of strike-slip faults is essential.

Methods A prediction model for strain peaks of pipelines under the action of strike-slip faults was developed based on the Separable Residual Neural Network (S-ResNet) model. To balance computational accuracy and efficiency, a parametric finite-element model was established by coupling pipe and shell elements. Mechanical responses under various factors—including pipe diameter, wall thickness, operating pressure, fault displacement, and soil parameters—were calculated, and a database of strain peaks was established. The separable self-attention mechanism was integrated into the residual neural network to construct the S-ResNet prediction model. Global and local interpretability analysis of the model was conducted using the SHapley Additive exPlanation (SHAP) method, identifying key factors influencing the prediction of tensile and compressive strain peaks.

Results Using finite element model results under various conditions as references, the S-ResNet–based strain peak prediction model for pipelines under the action of strike-slip faults demonstrated superior accuracy compared to common machine learning models. For tensile strain peak prediction, the model achieved a testing R² of 0.999 61, a mean square error (MSE) of 3.56×10⁻⁸, and a Mean Absolute Percentage Error (MAPE) of 1.751％. For compressive strain peak prediction, the model achieved a testing R² of 0.998 72, a MSE of 2.65×10⁻⁷, and a MAPE of 3.225％. In terms of interpretability analysis, SHAP analysis revealed that the strain peak was primarily influenced by strike-slip fault displacement, lateral soil spring ultimate resistance, pipe wall thickness, and crossing angle.

Conclusion The strain peak prediction model based on S-ResNet and SHAP enables rapid, accurate and interpretable prediction of pipeline mechanical responses under the action of strike-slip faults. It provides an efficient tool for pipeline safety evaluation in fault zones and serves as a valuable reference for digital twin development in geohazard-affected pipeline sections.