Class-imbalanced pattern recognition in pipeline weld cracks damage via feature characterization and sample enhancement

Acoustic emission (AE) technology can monitor the crack damage process of nuclear power pipelines, but the problem of class imbalance of signal and the difficulty of classical features to support the online assessment of damage evolution exist. For this reason, a method for online identification of pipeline weld crack damage patterns through feature characterization and sample enhancement is proposed in this paper. The proposed Accumulated Singular Value Energy Proportion (ASVEP) feature can quantitatively characterize the damage contribution and damage rate of the crack expansion process. The results show that the damage contribution and damage rate are maximized in the yield and strengthening stages of the crack expansion process.In addition, a well-designed Multilevel feature Fusion Omni-Scale convolutional neural network (MFOSCNN) can effectively recognize the damage patterns of a few types of samples in the crack expansion process. Compared with other state-of-the-art methods, the proposed online monitoring method has high recognition accuracy.