Abstract: The vibration monitoring data of the shearer cutting unit gearbox has a complex structure and is prone to class imbalance issues, leading to frequent false positives in traditional machine learning-based fault diagnosis methods. Meanwhile, deep learning-based approaches often suffer from complex model structures, low learning efficiency, and susceptibility to local optima, negatively impacting diagnostic performance. To address these issues, a fault diagnosis method was proposed for the shearer cutting unit gearbox based on improved cascaded broad learning (ICBL). A random hypergraph convolution mechanism was introduced into the feature nodes of the ICBL model to fully exploit the complex multivariate structural information in the vibration data of the shearer cutting unit gearbox, thereby enhancing the representation of fault features. Additionally, a class-specific weight allocation strategy was adopted to assign higher weights to minority class samples based on the class distribution of the input data, improving fault diagnosis performance under imbalanced data conditions. The effectiveness of the ICBL-based fault diagnosis method was validated using a shearer cutting unit gearbox fault simulation test platform. Experimental results demonstrated that the proposed method effectively enhanced the discriminability of fault features, achieving a diagnostic accuracy of 94.52% when the data imbalance ratio was 15, with a fault recognition time of 0.284 ms per sample. The method outperformed cascaded broad learning systems, weighted broad learning systems, multi-scale convolutional neural networks, hypergraph neural networks, and multi-resolution hypergraph convolutional networks, demonstrating significant engineering application value.