PLI and CNN-BiLSTM: An Enhanced Data Augmentation and Deep Learning Approach for Defect Recognition in Spiral Welded Pipe Based on Ultrasonic Guided Waves

Defect detection in urban heat pipes is critical for ensuring system safety and reliability, where Ultrasonic Guided Wave (UGW) technology plays a pivotal role. However, the complex guided wave propagation caused by spiral welds (resulting in low SNR) and the scarcity of valid defect samples (due to high acquisition costs) significantly reduce the performance of conventional deep learning algorithms. To address these challenges, we proposed a Proportional Linear Integration (PLI) model for data augmentation that preserves label invariance through linear transformations, accounting for the inherent linearity of UGW defect echo signals—a property often overlooked. PLI integrates four single-sample and two multi-sample linear transformation methods in equal proportions. Furthermore, we propose a hybrid CNN-BiLSTM architecture integrating Convolutional Neural Networks (CNNs) for local feature extraction with Bi-directional Long Short-Term Memory (BiLSTM) networks for long-range dependency capture, followed by LSTM to preserve the physical temporal sequence of UGWs. Using a small dataset of artificial defects (validated as equivalent to real defects), we applied original and PLI-augmented data to train our CNN-BiLSTM model, achieving defect localization error of only 0.2384 meters—outperforming other data augmentation methods and deep learning models. Our results highlight the method’s ability to significantly enhance defect localization precision in spiral-welded pipes, even with scarce samples. This study underscores the potential of tailored data augmentation and hybrid deep learning models for Non-Destructive Testing (NDT) and Structural Health Monitoring (SHM) of UGW time-series signals.