A Recognition Method for Ultrasound Signals of Spatiotemporal Nonlinear Distortion Using Adaptive Denoising and Domain-Adversarial Graph Convolutional Networks

Assessing the physical integrity of oil and gas wellbores is crucial for preventing fluid leakage and ensuring energy security. The ultrasonic pulse-echo method is a mainstream technology for detecting the cementing status outside the casing. However, in actual logging operations, physical environmental fluctuations such as transducer eccentricity and changes in the impedance of the internal fluid cause nonlinear distortion of the received ultrasonic signals. Meanwhile, the multiple reflection reverberations of the casing wall often mask weak interface defect features, leading to a significant decline in the cross-condition generalization ability of traditional deep learning models based on the independent and identically distributed assumption. To address these issues, this paper proposes a Domain-Adversarial Graph-Enhanced Denoising Network (DA-GEDN). This method constructs a 1D multi-scale residual encoder to adaptively extract local time-domain echo features. On this basis, the time-series signals are mapped into a topological graph structure, and a Graph Convolutional Network (GCN) is utilized to capture the spatio-temporal dependencies of the long-distance echo decay. To overcome the domain shift caused by changing operating conditions, an unsupervised domain adaptation mechanism based on a Gradient Reversal Layer (GRL) is introduced. Through a minimax game, it aligns the feature distributions of the calibrated conditions (source domain) and the complex testing conditions (target domain). Extensive experiments based on COMSOL ultrasonic simulations demonstrate that, when facing combined interferences such as transducer eccentricity and fluid replacement, DA-GEDN can effectively extract domain-invariant acoustic features. Its recognition accuracy for micro-gaps and debonding defects outside the pipe reaches 83.42%, significantly outperforming traditional baseline models. Additionally, its macro-F1 score in the ablation study is higher than that of all other comparative models.