<p>In RIS-assisted multi-user uplink MIMO systems, existing deep learning (DL)-based channel estimation methods generally depend on clean channel labels that are difficult to obtain in practice and exhibit limited adaptability to time-varying noise conditions. To address these issues, this paper proposes a unified noise-aware multi-task self-supervised learning (MT-SSL) framework for RIS-assisted uplink channel estimation. In the training phase, the proposed framework adopts a self-supervised denoising formulation with internally constructed noisy-clean training pairs, while an auxiliary noise-intensity prediction task is introduced to guide adaptive feature fusion according to the estimated noise condition of each input sample. In the testing phase, the least-squares (LS) channel estimate is used as the input, and the trained model refines the coarse LS estimate to recover the cascaded channel more accurately. In this way, the proposed method reduces the dependence on perfect CSI supervision while improving robustness under time-varying noise conditions. Simulation results show that, compared with the conventional self-supervised learning (SSL) algorithm, the proposed method reduces the normalized mean square error (NMSE) by 15.05%. Moreover, the results demonstrate that the joint use of noise-aware adaptive fusion and multi-task learning improves the adaptability and robustness of the model in self-supervised channel refinement, with the most evident gain observed within an intermediate range of the injected-noise standard deviation.</p>