Abstract: As the primary equipment for roadway excavation, the roadheader operates under highly complex and variable working conditions, which makes the cutting head prone to faults. However, harsh excavation environments cause the cutting signals of the roadheader to exhibit strong background noise coupling and pronounced non-stationary characteristics, which significantly hinder fault feature extraction and lead to low diagnostic accuracy. To address these challenges, a Frequency Dynamic-aware Multi-view Tree-topology Diagnosis Network (FDM-TreeDNet) is proposed. In this network, a novel Frequency Dynamic-Aware Convolution (FDAC) module is designed. The module exploits the spectral centroid of channels to guide channel reordering and, by leveraging the energy distribution characteristics of high- and low-frequency signals, employs anisotropic convolutions to accurately decouple high- and low-frequency features. Subsequently, a multi-branch convolutional gating mechanism is adopted to capture real-time semantic information and dynamically generate sample-specific weights. While achieving frequency awareness, the FDAC module also performs adaptive feature reconstruction and calibration, thereby suppressing strong background noise and effectively alleviating feature drift under non-stationary operating conditions with severe noise interference.To further address the issue of semantic degradation during deep feature propagation, a Source-Guided Multi-view Tree-topology (SGM-Tree) structure is constructed. In this structure, lateral injection paths are established at key stages of the backbone network, where parallel branches explicitly decompose multi-view critical features from the raw signal, including precise positional information, long-range structural representations, and global contextual semantics. These features are progressively injected as semantic anchors prior to the FDAC modules, effectively correcting semantic deviations introduced by frequency reordering. To fully exploit the hierarchical feature advantages brought by the tree-topology architecture, a Tree-Structured Distillation (TSD) strategy is further introduced. Through a parameter-sharing classification mechanism, the multi-view critical features injected by the SGM-Tree at different stages and the dynamic output features of the backbone network are projected into a unified semantic metric space. Within this space, confidence transformation is employed to obtain semantic soft masks, which, combined with multi-level cascaded self-distillation, enhance the fault diagnosis accuracy of the roadheader cutting head. Experimental results demonstrate that the proposed roadheader cutting head fault diagnosis approach achieves superior performance on both the self-constructed AUST roadheader dataset and the CWRU dataset, thereby validating its effectiveness.