Abstract: To address unstable oil–water layer identification in the S-well logging series of the N block, Daqing Oilfield, caused by thin interbeds, noise, and missing intervals, we propose a long-sequence logging modeling approach integrating a State Space Model (SSM) for cross-well generalization. Seven log channels are preprocessed with despiking/detrending, log-domain transformation, missing-value completion with masks, and standardization. A well-level split is adopted to build training/validation/test sets (23 wells, 85, 600 depth points at Δ=0.125 m; oil proportion ≈34%). The model employs multi-scale 1D convolutions to capture local bed-scale patterns, an SSM linear-time scan to aggregate long-range context, and gated fusion with total-variation (TV) continuity regularization to suppress depth-wise jitter, enabling point-wise oil/water classification. On the test set, the proposed method outperforms Archie, Random Forest, and 1D-CNN, achieving Acc=0.91, F1=0.89, AUC=0.94, and MCC=0.78 with an inference time of 7.5 ms per interval. Ablation studies and depth-aligned visualization confirm that long-range modeling and continuity constraints are crucial for stable boundary delineation in thin interbedded sections.