Acoustic Emission Localization of Welded Pipes Using XGBoost Based on SHAP Interpretation

Abstract In the structural health monitoring of pipelines, the accurate localization of acoustic emission (AE) signals generated by crack propagation is critical for leakage prevention. The accuracy of the traditional time difference of arrival (TDoA) localization method drops sharply due to waveform distortion, which severely restricts the engineering application of acoustic emission technology. To address this challenge, this paper proposes an interpretable artificial intelligence localization method validated against physical principles. This method employs an XGBoost model optimized via Bayesian optimization (BO), which avoids reliance on constant wave velocity by learning multi-dimensional acoustic features. Experimental results demonstrate that the proposed model significantly outperforms the traditional TDoA method, reducing the mean absolute error (MAE) and root mean square error (RMSE) by 71.7 and 77.4%, respectively. The core contribution of this work lies in revealing the intrinsic mechanism behind the model’s success through SHAP (SHapley Additive exPlanations) interpretability analysis: it achieves a transformation of the localization paradigm – independently discarding the time difference information that fails at welds, and instead learning and leveraging the acoustic low-pass filtering effect of welds, thereby effectively converting this “structural defect” (in the traditional sense) into a high-precision “acoustic landmark”. This study not only provides an intelligent localization solution with high accuracy and strong robustness, but also facilitates the in-depth application of reliable data-driven methods in the field of structural health monitoring by quantitatively verifying that the model’s decisions are consistent with physical laws. The model’s efficacy is validated under the specific experimental conditions of a straight pipe with a single butt weld, for burst-type AE sources located along a single axial path.