Acoustic-vibration dual-level fusion for leakage quantification in pipeline flanges

Multi-sensor leakage detection based on a single physical characteristic of gas leakage lacks complementarity and exhibits poor adaptability to complex leakage conditions. Although multi-source information is complementary, direct fusion may suffer from information misalignment and redundancy and may introduce additional noise. This study proposes a pipeline flange leakage detection method that combines low-level feature fusion with high-level model integration. Time-frequency features of leakage-induced acoustic and vibration at the flange are separately extracted. To address data redundancy, a Pearson correlation coefficient combined with a random forest algorithm is employed to reduce the dimensionality of the noise and vibration feature vectors. A cross-attention mechanism is applied to perform weighted fusion of the dimension-reduced features. The fused features are then fed into eXtreme Gradient Boosting (XGBoost) and Light Gradient Boosting Machine (LightGBM) models, with the leakage rate as the prediction target, to obtain individual prediction results. An entropy weight method is introduced to construct a parallel fusion framework for XGBoost and LightGBM, enabling collaborative regression prediction of the leakage rates. Experimental results demonstrate that the model trained with this method achieves a root mean squared error of 1.363 and an R 2 of 0.9931. Compared to using individual vibration or acoustic feature matrices as input, feature fusion improved the R 2 of XGBoost by 2.66 and 1.53%, and that of LightGBM by 2.98 and 2.16%, respectively. Furthermore, compared to using only XGBoost or LightGBM individually, the fusion model achieves relative R 2 improvements of 0.6 and 0.8%. The proposed method in this study, through multi-source feature fusion and model synergy, effectively enhances both the accuracy and robustness of leakage rate prediction.