Deep Learning Model and Application for 3D Morphology Reconstruction of Casing Damage

This paper addresses the core issues of data sparsity, irregular morphology, and noise interference in the 3D morphology reconstruction of casing damage. A multi-scale feature fusion Transformer (MSF-Transformer) deep learning model is proposed. The model adopts a hierarchical architecture of “preprocessing - feature extraction - damage enhancement - feature fusion - reconstruction output.” Through a multi-scale Transformer encoder constrained by local neighborhoods, it simultaneously captures the macroscopic structure and fine details of the damage, reducing the computational complexity from O(N²) to O(N×K). A damage region enhancement module (DREM) is designed to generate a dynamic mask based on feature similarity, enhancing the weak feature response of minor damages and suppressing noise interference. A multi-objective hybrid loss function is constructed, fusing normal-guided chamfer distance, normal vector, and volume loss to balance morphological consistency and quantization accuracy. Experiments were conducted using 800 dedicated datasets covering different damage types, scales, and noise levels. Results show that the MSF-Transformer achieves a MAE of 0.032 mm and an RMSE of 0.047 mm in a noise-free environment, representing reductions of 18.9% and 21.7% respectively compared to the TPR. For 0.5 mm microcracks, the RVRE is only 4.2%, and the MAE is 0.071 mm under 20% high-intensity noise, exhibiting the smallest performance degradation. The single-sample reconstruction time is 23.6 ms, and the GPU memory usage is 1.8 GB, meeting real-time detection requirements. This model achieves high-precision, noise-resistant, and high-efficiency 3D reconstruction of casing damage, providing reliable technical support for casing integrity assessment.