针对含油饱和度解释依赖岩心分析与经验公式、标注样本稀缺且测井曲线缺失与噪声显著的问题，提出一种基于掩码自监督Transformer的测井含油饱和度预测方法——MWLT-So。该方法首先通过深度对齐与标准化对多种常规测井曲线预处理，并构建随机掩码重建任务，在大量无标签井段上进行自监督预训练，以提取具有跨井迁移能力的通用表征。随后，在少量标注样本上微调模型，结合多尺度位置编码与特征融合机制，有效提升了对长程依赖关系的建模精度。跨区块与不同测井组合的对比验证表明：①在大庆油田N区块测井数据集与按井集划分的实验设置下，MWLT-So在回归主任务上取得最优性能。②MWLT-So在地质剖面一致性与误差分布上均具优势。该方法能精准刻画层界过渡带与高S_o平台段，有效克服过平滑、过冲及相位滞后等传统缺陷；其残差分布更集中、尾部更薄，误差中位数与离散度均更低。这表明模型兼具高精度与高稳健性。③基于阈值t = 0.5的油层/非油层识别任务中，MWLT-So以最低的假阳性(FP)与假阴性(FN)取得最优分类性能，验证了其在阈值分层场景下降低误判风险的能力。该方法可为复杂储层快速评价与剩余油识别提供支撑。

Accurate interpretation of oil saturation (S_o) traditionally relies on core analysis and empirical formulas, yet this process is challenged by scarce labeled samples, missing logging curves, and significant noise in logging data. To address these issues, this paper proposes a masked self-supervised Transformer-based method for oil saturation prediction from well logs, termed MWLT-So. First, multiple conventional logging curves are thoroughly aligned and standardized, and a random masking reconstruction task is designed to enable self-supervised pretraining on a large amount of unlabeled well interval, thereby learning generalizable representations with cross-well transferability. The pretrained model is then fine-tuned on a limited number of labeled samples using a regression objective, while multi-scale positional encoding and feature fusion are incorporated to enhance long-range dependency modeling. Comparative experiments across different blocks and various logging-curve combinations demonstrate that:(1) Under both random data split and rigorous well-wise split evaluation settings, MWLT-So achieves the best performance on the primary regression task. (2) MWLT-So shows clear advantages in terms of geological profile consistency and error distribution, accurately capturing boundary transition zones and high-S_o plateau intervals, while effectively overcoming common deficiencies of traditional methods such as over-smoothing, overshooting, and phase lag. Its residuals are more concentrated with thinner tails, and both the median error and dispersion are lower, indicating superior accuracy and robustness. (3) In the oil-bearing/non-oil-bearing identification task based on the threshold S_o = 0.5, MWLT-So attains the best classification performance with the lowest false-positive and false-negative rates, confirming its capability to reduce misclassification risk in threshold-based stratification scenarios. Overall, the proposed method provides effective technical support for rapid evaluation of complex reservoirs and identification of remaining oil.