Concealed collapse columns, as typical hidden geological anomalies in coalfields, directly impact coal mine safety and geological hazard prevention through the accuracy of their boundary identification. To address the limitations of traditional single-attribute methods in responding to weak boundaries and suppressing noise, this paper proposes a multi-scale characterization and enhancement method that integrates structure-oriented filtering and frequency-divided coherence attributes. Based on 3D seismic data from a mining area in Shanxi, the structure-oriented filtering technique is first applied, combining gradient structure tensors and anisotropic diffusion equations to effectively suppress random noise while significantly preserving the steeply dipping structural features of collapse column boundaries. Subsequently, short-time Fourier transform is used to perform spectral decomposition on the filtered data, extracting amplitude and phase attributes of multiple single frequencies within the 40–100 Hz range. This systematically reveals the frequency-dependent characteristics of seismic responses at collapse column boundaries: low-frequency (60–70 Hz) amplitude attributes provide good indications for large-scale collapse column outlines, while high-frequency (80–90 Hz) phase attributes exhibit superior sharpening and resolution capabilities for small-scale collapse column boundaries. Furthermore, the eigenvalue coherence algorithm is introduced to quantify formation discontinuities, and a multi-frequency attribute fusion strategy is employed to achieve integrated enhancement and fine characterization of collapse column boundaries in spatial distribution. Practical data applications demonstrate that this method significantly improves the signal-to-noise ratio of seismic data and the accuracy of boundary identification, providing a reliable multi-scale geophysical technique for the detection and interpretation of concealed collapse columns in coalfields.