Modern radar signal sorting faces critical challenges in complex electromagnetic environments characterized by severe parameter overlapping and dynamic modulation patterns. Traditional methods relying on single-domain parameter analysis (e.g., RF, PW, DOA) exhibit limited robustness under such conditions, while existing PRI-based approaches struggle with multi-modulation adaptability and computational efficiency. This paper proposes a self-organized radar signal sorting framework integrating multi-order DTOA spectral analysis and parameter entropy dynamic fusion. First, frequency-domain features are extracted from TOA sequences through multi-order DTOA spectral analysis, where threshold-controlled DTOA frequency matrices are transformed into latent PRI features via softmax-based correlation mapping. Second, an adaptive weighting mechanism dynamically fuses multi-domain features by evaluating feature entropy and standard deviation, generating a weighted similarity matrix that captures inter-pulse correlations. Third, manifold learning optimizes the high-dimensional similarity matrix into a low-dimensional discriminative space for clustering. Experiments on five simulated radar pulse descriptor datasets demonstrate the framework’s superiority over state-of-the-art methods, achieving average Rand Index (RI) of 0.960 and Adjusted Rand Index (ARI) of 0.902. The proposed method resolves parameter overlap, uneven pulse density, and PRI modulation diversity in modern radar signal sorting for congested environments.