| description abstract | Abstract. Offshore platform pipeline leakage detection faces severe challenges from complex marine environments, where intense environmental noise interference and complex signal characteristics make traditional methods difficult to achieve accurate leakage valve localization. To address this technical challenge, this study proposes an offshore platform pipeline leakage valve localization method based on dynamic time warping distance-based complete ensemble empirical mode decomposition with adaptive noise (DCEEMDAN) and adaptive temporal–spatial fusion network (ATSFN). First, by introducing dynamic time warping distance similarity measurement into the complete ensemble empirical mode decomposition with adaptive noise (CEEMDAN) framework and combining probability density function feature extraction, adaptive denoising of acoustic emission signals in marine environments is achieved. Second, a temporal–spatial feature extraction architecture with a parallel multiscale convolutional neural network (CNN) and a hierarchical GRU is designed, realizing deep fusion of CNN spatial features and GRU temporal features through a cross-attention mechanism. Finally, an end-to-end intelligent monitoring system is constructed, achieving high-precision localization of 10 valve positions through dual-stage verification combining laboratory experiments and offshore platform field measurements. Experimental results show that DCEEMDAN outperforms traditional EMD series algorithms, achieving a signal-to-noise ratio (SNR) of 19.69 dB with 16.6% improvement over CEEMDAN. ATSFN achieves average localization accuracies of 94.38% and 95.75% under 4 MPa and 5 MPa conditions, respectively, representing improvements of 10.9% and 10.77% over best baseline models. Under extreme noise conditions, the model maintains localization accuracy above 82.3%, demonstrating excellent noise robustness. This research provides an effective technical solution for offshore platform pipeline leakage detection. | |
