Abstract: Well logs from the S-series wells in Block N of the Daqing Oilfield are affected by thin interbeds, random noise, abnormal spikes, baseline drift, and continuous missing intervals, which can lead to overlapping oil–water responses, missed detection of thin oil layers, boundary shifts, and fragmented predictions along the depth direction. To address these issues, a long-sequence well-log oil–water layer identification method integrating a state-space model (SSM) is proposed. Seven-channel well logs are first processed by despiking, detrending, logarithmic transformation, missing-value imputation, and standardization, and are then divided into training, validation, and test sets on a well-by-well basis to prevent information leakage from the same well. Multi-scale one-dimensional convolutions are subsequently employed to extract local responses from stratigraphic intervals of different thicknesses. The SSM performs linear scanning to model stratigraphic dependencies and the evolution of fluid responses over extended depth intervals, while a gated fusion mechanism coordinates local thin-layer features with long-range interval information. Total variation (TV) regularization is further introduced to constrain unnecessary variations in prediction probabilities between adjacent depth samples, thereby improving consistency with vertical stratigraphic continuity. The proposed method achieves an accuracy of 0.91, an F1 score of 0.89, an AUC of 0.94, and an MCC of 0.78. The correct identification rates for oil and water layers are approximately 88.0% and 92.5%, respectively, and the overall performance is superior to those of the Archie equation, random forest, and one-dimensional convolutional neural network. Its inference time is 7.5 ms, lower than the 9.8 ms of the 1D-CNN. The proposed method effectively reduces missed classifications, boundary shifts, and fragmented predictions under complex thin-interbedded conditions, thereby improving the continuity, stability, and interpretational reliability of oil–water layer identification.