<p>Cutting head is a part of the roadheader prone to failure. Its health monitoring and fault diagnosis can ensure the safe and efficient operation of the roadheader. The cutting head has been under the complex working condition of variable load for a long time. Therefore, the vibration signal of the cutting head has non-linear time-varying modulation characteristics, which causes serious interference to the fault identification of the cutting head. Hence, this study proposes a feature extraction method based on rime algorithm (RIME) optimized variational mode decomposition (VMD) and refined composite multi-scale fluctuation dispersion entropy (RCMFDE). Meanwhile, it employs the advantages of Deep Belief Networks (DBN) in nonlinear high-dimensional data processing to classify and recognize the failure modes of the cutting head. Firstly, the paper obtains the optimal parameter combinations of the VMD algorithm through the RIME algorithm. It uses the optimized VMD to adaptively decompose the cutting vibration signal and get a series of intrinsic modal functions (IMF). The paper combined the correlation coefficients to screen the optimal eigencomponent. Then, for the feature IMF component, this study explores the impact of the embedding dimension and category number of RCMFDE on the feature extraction performance. It calculates the RCMFDE of vibration signals from different cutting heads and uses them as the eigenvector. Finally, the paper uses the DBN model to train and test the cutting vibration features and realize the fault pattern recognition of the cutting head. The simulation and experimental results show that the proposed method can effectively extract the fault characteristics of cutting vibration signals, and the recognition accuracy reaches 99.164%. Compared with other methods, it has better recognition accuracy and robustness and can provide a new research idea for monitoring the health status of the cutting head.</p>