To address the challenges of low signal-to-noise ratio （SNR） and the lack of clean ground truth labels for distributed acoustic sensing （DAS） and distributed temperature sensing （DTS） data during hydraulic fracturing，a physics-informed self-supervised blind denoising method is proposed. Instead of relying on clean reference data，this method constructs self-supervised training samples using a checkerboard alternating sampling strategy on time-depth sub-blocks and enables the denoising network to learn noise distributions adaptively. Furthermore，physical constraints such as thermal conduction smoothness，edge preservation，and wavefield coherence are introduced for DTS and DAS characteristics，respectively. A spatially weighted loss function based on perforation cluster positions is also established to prioritize the preservation of key response features near clusters. It is demonstrated that the proposed method outperforms traditional filtering methods in terms of SNR and structural similarity （SSIM）. The denoised data significantly enhances the clarity of cooling fronts and acoustic energy bands at fluid entry points，which effectively improves the accuracy of fracturing event recognition and cluster efficiency evaluation.