Offshore oil and gas pipelines face severe corrosion, fatigue, and leakage risks due to extreme marine environments and multi-stress coupling. Traditional acoustic emission (AE) monitoring techniques struggle with significant background noise and high-dimensional feature redundancy, which reduces damage identification accuracy. In this study, we propose a detection framework that combines adaptive sliding window feature extraction with a joint analysis of covariance and correlation matrices. To avoid the subjectivity of conventional thresholding, we use an adaptive sliding window to construct a 50-dimensional time–frequency feature space from AE signals collected during hydrostatic loading experiments on prefabricated cracked pipelines. By jointly analyzing the covariance matrix and Pearson correlation coefficients, we select three low-redundancy, high-sensitivity complementary features (count rate, average frequency, and waveform complexity) from 14 candidates. This achieves a dimensional compression rate of 78.6%. Ultimately, the overall compression ratio of the feature space for each channel reached 68.0%. We construct an artificial neural network with three hidden layers, incorporating a focal loss function and K -fold cross-validation to handle sample imbalance. This enables precise classification of pipeline damage across four stages: elastic deformation, crack initiation, plastic propagation, and fracture failure. The model achieves 93% accuracy on an independent test set and 87% on a new working condition validation set. Experimental results reveal a strong correlation between AE signals and mechanical behaviors. Specifically, the elastic phase exhibits low-amplitude harmonics, the plastic phase shows intermittent jumps, and the fracture phase features the release of high-frequency energy. Ultimately, this method provides a reliable diagnostic solution for the structural health monitoring of offshore pipelines. Coupling efficient feature selection with robust neural network modeling significantly improves state identification and traceability of dynamic crack propagation.
Pan et al. (Fri,) studied this question.