Wellbore stability is a key factor affecting safe and efficient drilling. At present, it is difficult to conduct real-time and accurate analysis of wellbore stability in related research. To address the current research shortcomings, this study proposes a real-time analysis model of wellbore stability integrating image recognition and an expert system, which mainly includes caving image segmentation and recognition, and a wellbore stability expert system. The caving image recognition proposes a new dynamic threshold segmentation method based on simple linear iterative clustering superpixel segmentation and visual geometry group 19-layer image classification. After completing the segmentation of the caving image, the geometric features of the caving are calculated, and the multi-source feature fusion GoogleNet model is established by integrating the geometric features with the convolution features extracted by GoogleNet to identify the caving types efficiently. After segmentation and recognition of caving images. The wellbore stability expert system uses the caving features to establish an expert system model to determine the mechanism of wellbore instability and provide reasonable solutions. Finally, the wellbore stability integrating image recognition and an expert system model was applied to a well in field production, accurately determining the mechanism of wellbore instability in real time and effectively solving the corresponding wellbore instability problem based on the measures provided by the model. Document Type: Original article Cited as: Fan, Y., Pang, H., Jin, Y., Meng, H., Lu, Y., Wei, S., Wang, H. Integration of image recognition and expert system for real-time wellbore stability analysis. Advances in Geo-Energy Research, 2025, 15(2): 158-171. https://doi.org/10.46690/ager.2025.02.07

The morphological evolution of mineral-filled fractures in low-clay shale during water exposure remains poorly understood, particularly regarding how infilling mineralogy, fracture geometry, and fluid accessibility control the response. In this study, fixed-field scanning electron microscopy coupled with energy-dispersive X-ray spectroscopy (SEM–EDS) was used to track selected fracture regions in Longmaxi Formation shale recovered from a depth of approximately 4072 m in the southern Sichuan Basin, China. The same regions were examined before immersion and after 3 and 6 days of static immersion in deionized water (18.2 MΩ·cm; essentially zero initial ionic strength). Changes in fracture trace length, apparent aperture, area fraction, and box-counting fractal dimension were quantified using a consistent image-analysis procedure and interpreted together with elemental distributions. Three fracture-evolution patterns were identified: sustained enlargement, initial enlargement followed by partial reduction, and moderate aperture increase accompanied by pronounced trace-length extension. Calcium-sulfate-rich fractures exhibited the first two patterns. Their initial enlargement was associated with dissolution of the mineral infilling, whereas the contrasting later-stage responses appeared to be influenced by fracture geometry and fluid accessibility. Relatively open fractures continued to enlarge, while more confined fractures showed partial loss of the initial enlargement and local accumulation of crystalline material. However, the composition and origin of this material (including possible drying-induced crystallization) remain uncertain. Clay-rich fractures showed modest aperture enlargement and continued trace-length extension, consistent with clay–water interaction, although the underlying mechanism could not be resolved conclusively. These results demonstrate that fractures with similar mineral infillings can follow different temporal evolution paths during water exposure. The fixed-field SEM–EDS workflow provides a reproducible method for quantitatively tracking microscale fracture evolution and improves understanding of mineral- and geometry-dependent fracture responses in low-clay shale, though the results are derived from a single specimen and should be interpreted as illustrative observations rather than population-level statistics.