Abstract: To address the insufficient cross-well generalization of oil layer identification models caused by strong noise in well logging curves and pronounced inter-well distribution differences, an end-to-end reservoir identification method integrating TS-MAE self-supervised pre-training with the variable-wise attention mechanism of iTransformer is proposed. First, multichannel well-logging data are subjected to alignment, missing-value imputation, and standardization. Oil layer, water layer, and dry layer samples are then constructed using sliding-window segmentation with a window length of 64 and a stride of 16, together with center-point labeling. Second, a masking ratio of 0.75 is adopted during pre-training stage, and TS-MAE is employed to reconstruct the well-logging data and learn transferable joint representations of multiple logging curves. Finally, the pre-trained encoder is transferred to the downstream iTransformer classification network for fine-tuning. Experimental results based on 3,705 samples from four wells show that the proposed method achieves an accuracy of 93.7%, a Macro-F₁ score of 93.1%, and an AUC of 0.971 on the test set, outperforming traditional formula-based methods, conventional machine-learning baselines, and an iTransformer trained from scratch. In cross-well validation, the model achieves a Macro-F₁ score of 89.5% and an AUC of 0.956, demonstrating strong robustness to distribution shifts in unseen wells. Ablation experiments and visualization analyses further confirm the synergistic contributions of self-supervised pre-training and variable-wise attention to performance improvement, while enhancing the physical consistency and interpretability of the model predictions. The proposed method provides an efficient and deployable intelligent solution for multi-well reservoir evaluation.