Abstract:Addressing the limitations of rock tensile strength test, such as high equipment requirements and complex sample preparation, rock density, specimen diameter, Young"s modulus, static tensile strength, loading rate and test method were selected as prediction indicators, with dynamic tensile strength as the output indicator. A sample database for predicting rock dynamic tensile strength was constructed with 164 sets of sample data. By combining the long sequence modeling ability of Temporal Convolutional Networks (TCN) and the interpretable nonlinear transformation advantage of Kolmogorov-Arnold Networks (KAN), a TCN-KAN combined model for predicting the dynamic tensile strength of rock was constructed. The model was trained by five-fold cross validation, and the interpretability of the model prediction results was analyzed using the SHAP method. The results showed that the combined model outperformed the comparative models in terms of mean square error (3.07), root mean square error (1.75), mean absolute error (1.27), mean absolute percentage error (10.22%) and determination coefficient (97.88%). The loading rate and static tensile strength had the most significant impact on the prediction results. Finally, based on the TCN-KAN combined model, an intelligent application software was developed and applied in five engineering cases, further verifying the predictive accuracy and reliability of the combined model. This provides an intelligent new method for predicting rock dynamic tensile strength.