In industrial plants, pipelines are widely used to transport various fluids, and internal blockages caused by deposit accumulation can significantly compromise equipment integrity. Accurate identification of the presence and location of such blockages is therefore essential. Conventional diagnostic methods rely on visual inspection and radiographic testing (RT); however, visual inspection requires pipe disassembly, while RT suffers from limited applicability and safety concerns associated with radiation exposure. In this context, digital hammering inspection has emerged as a safe and non-destructive alternative for pipe blockage diagnosis[1]. This technique measures vibration waveforms induced by tapping and can be performed easily in situ. Nevertheless, current diagnostic criteria remain largely subjective, highlighting the need for quantitative and objective evaluation methods. Meanwhile, machine learning has been increasingly applied in structural health monitoring to infer structural conditions from measured signals. Its ability to capture nonlinear relationships makes it well suited for analyzing complex vibration responses. Among various algorithms, XGBoost, a tree-based ensemble learning method, offers high predictive accuracy along with feature importance measures that enhance model interpretability. In this study, a machine learning–based framework for pipe blockage diagnosis is proposed. A piping model with simulated internal blockages was analyzed using the dynamic finite element solver LS-DYNA, and vibration responses generated by hammering were obtained. Acoustic features were extracted using formant analysis and Mel-frequency cepstral coefficients (MFCCs), and a surrogate classification model was constructed using XGBoost[2][3]. The proposed method was validated by classifying the presence, location, and severity of pipe blockages, demonstrating its effectiveness for quantitative and non-destructive diagnosis.