In response to the challenges of unknown seismic noise distribution and difficulties in constructing paired samples, we propose a self-supervised blind denoising method integrating dynamic noise modeling with collaborative optimization. The method constitutes a dual-network collaborative architecture comprising a noise modeling subnetwork and a denoising subnetwork. Both subnetworks adopt U-shaped network structures incorporating wavelet transforms and Swin Transformers. The noise modeling subnetwork is optimized through mean squared error, while the denoising subnetwork is trained using a SURE-based loss function. The dual-network collaborative framework overcomes limitations of traditional supervised learning by: (1) introducing a non-prior modeling strategy that dynamically estimates noise distribution characteristics through data-driven noise modeling sub-network; and (2) sampling predicted noise distributions to construct nosier-noisy sample pairs for training the denoising subnetwork. The two subnetworks undergo collaborative training with mutual promotion and synchronous convergence, achieving blind noise modeling and blind denoising respectively. Numerical simulations and practical seismic data experiments demonstrate the proposed algorithm's effectiveness.