Background Accurate identification of the dense weathered zone at the top of the Ordovician strata is essential for mine-water hazard prevention and control in North China coalfields. However, for this specific type of weathered zone, the applicability of existing identification methods remains limited, mainly due to insufficient calibration data, ambiguous logging responses, and the subjectivity of conventional empirical interpretation.

Methods Based on the stratigraphic occurrence patterns and hydrogeological evolution characteristics of the weathered zone at the Ordovician top, a three-stage stepwise identification framework was established. The framework consists of geological-prior delineation of the top boundary, logging-response interpretation of hydrogeological characteristics, and verification of the bottom boundary. The combined responses of natural gamma ray (GR), deep lateral resistivity (LLD), spontaneous potential (SP), and bulk density (DEN), together with stratigraphic constraints, were clarified. On this basis, weathered-zone identification was formulated as an anomaly detection problem constrained by the unweathered bedrock background. A multi-scale convolutional, attention-enhanced, Transformer-integrated autoencoder model, termed MS-TransAE, was proposed. By adopting a “unweathered bedrock for training and shallow strata for anomaly identification” strategy, the model learns the logging patterns of unweathered bedrock and adaptively determines the anomaly threshold using extreme value theory.

Results and Conclusions  The proposed method was applied to 21 boreholes in the Dongpang Mine, Hebei Province. The identified thickness of the potential weathered zone at the Ordovician top ranges from 0 to 26.35 m, with an average thickness of 17.69 m. The absolute errors of the identified bottom boundaries are all less than 1.5 m. Ablation experiments show that multi-scale convolution, multi-scale attention, logging-curve morphological features, and the lithology classification head play important roles in improving reconstruction performance, threshold stability, and regularized learning. Comparative experiments further show that, compared with conventional unsupervised anomaly detection models, MS-TransAE reduces the average E_MA by 77.3%–83.0% and improves the average IoU by 20.7%–46.9%. Based on the identified weathered-zone horizons, the grouting horizons for regional grouting treatment in the Dongpang Coal Mine were calibrated. The calibrated grouting elevations range from −819.55 to −289.92 m, corresponding to depths of 0–69.04 m below the Ordovician top, with an average depth of 26.98 m. These results provide a reference for identifying dense weathered zones at the top of the Ordovician strata and for floor water-hazard control under similar geological conditions in North China coalfields.