A data augmentation method for pipeline weld defects based on LSTM-SSL-ACGAN

Pipeline weld defect detection is crucial for ensuring industrial production quality and pipeline operational safety. With the widespread application of intelligent detection technologies, high-performance deep learning models require large-scale, high-quality datasets for training. However, in practical working conditions, environmental magnetic interference and magnetic memory signal attenuation lead to challenges such as scarcity of high-quality samples, difficulty in feature extraction, and low recognition rates in small-sample scenarios. To address these issues, this study proposes a data augmentation method based on the LSTM-SSL-ACGAN framework. Its innovation lies in leveraging LSTM to capture deep features of magnetic signals and achieving distribution alignment through SSL-ACGAN constraints, combined with spectral normalisation to suppress noise interference. Experiments employing multi-dimensional data augmentation evaluation metrics and a CNN classification model validation framework demonstrate significant improvements in multi-class defect recognition accuracy, precision, recall, and F1-score. Specifically, precision for cracks, porosity, and undercut increased by 8.8%, 8.0%, and 13.0% respectively; recall improved by 8.4%, 7.5%, and 11.3%; and F1-scores rose by 8.6%, 7.8%, and 12.2%. Overall defect recognition accuracy increased by 9.0%. This method significantly enhances data quality and model robustness, providing a reliable solution for industrial pipeline weld defect detection.