Precise identification of pipeline surface defects utilizing ultrasonic guided wave technology and data-driven algorithmic approaches

Long-range ultrasonic guided waves (UGW) represent a critical approach for pipeline defect detection. However, the intricate propagation characteristics of high-frequency UGW signals, combined with the inherently low signal-to-noise ratio of weak defect-related features, present substantial challenges for effective detection and interpretation. To address these issues, this study proposes a novel multi-dimensional feature boosting method for UGW signals combined with a defect prediction framework, PKO-RF-Importance-CNN-BiLSTM, integrating the Pied Kingfisher Optimizer (PKO) with Random Forest (RF) feature importance analysis, Convolutional Neural Networks (CNN), and Bidirectional Long Short-Term Memory (BiLSTM) networks. Firstly, a feature selection method (PKO-RF-Importance) using PKO and Importance analysis to optimize RF is proposed, which overcomes the instability of feature extraction in small-sample learning, facilitates workload assessment, and constructs an optimal feature subset that not only enhances the spatio-temporal characterization of UGW signals, but also enhances the performance of the subsequent deep learning model. Secondly, a data-driven prediction framework (PKO-CNN-BiLSTM) is constructed by optimizing a combined CNN-BiLSTM network architecture using the PKO algorithm. This model significantly improves the learning capacity for nonlinear features and subtle weakly correlated features in UGW signals, thereby enhancing generalization ability. Experimental verification using UGW signals collected from 304 stainless steel pipelines with artificial flat-bottomed hole defects of varying sizes, depths, and transducer positions validates the efficacy and superior performance of the proposed methodology.