High-Accuracy Cement Bond Evaluation via Transfer Learning: An Integrated Framework for Microannulus Detection

Summary Convolutional neural networks (CNNs) have strong potential in automated cement evaluation, as they can detect subtle acoustic signatures linked to wellbore defects. However, three main challenges hinder their field application: limited data, biases in feature extraction, and blurred time-frequency (TF) representations from the conventional continuous wavelet transform (CWT). We propose VRNet, a transfer learning framework combining feature fusion and attention mechanisms for high-precision detection of microannuli at the casing-cement (C-C) interface. Key innovations are (1) enhanced TF analysis—replacing CWT with the wavelet synchrosqueezed transform (WSST) to reduce spectral blurring, yielding sharper patterns for the network to learn from. VRNet trained on WSST images achieved 98.7% accuracy vs. 95.5% with CWT, a +3.2% gain; (2) robust feature learning—integrating two complementary deep network architectures (ResNet-50 and VGG-16) through feature fusion and using an attention module to focus on the most informative TF regions; and (3) rigorous validation—trained on 768 simulated pulse-echo (PE) signals from casedhole wells and tested on experimental data from a scaled laboratory model with microannuli created by cyclic pressure-temperature treatment, achieving 98.7% accuracy on simulated data and 90.2% on experimental data. Gradient-weighted class activation mapping (Grad-CAM) shows that VRNet emphasizes the 175–225 kHz range, consistent with the physics of microannulus-induced signal reverberations. This capability supports more reliable well integrity assessments in operational settings, enabling earlier detection of microannuli and informing design improvements for ultrasonic logging tools in the field.