Distributed optical fiber vibration sensing (DOFS) monitors environmental vibrations by demodulating phase disturbances. Accurate detection of high-order vibration frequencies is essential for inferring complex physical events. To enhance the characterization of high-frequency modes in current systems, a feature extraction framework based on differential-manifold topology is proposed. First, the input signal is mapped into a high-dimensional phase-space trajectory via differential time-delay embedding, which nonlinearly reconstructs the dynamics and reveals latent high-order characteristics. Next, a simplicial structure is constructed using the Vietoris–Rips complex, and persistent homology is applied to quantify the robustness of topological features, yielding a topology-invariant persistence diagram. Finally, the Persistent Spectral Enhancement algorithm employs a convolutional network to optimize the discriminative representation of these topological features.