Abstract: As critical transmission components in industrial machinery, gearboxes require accurate fault diagnosis to ensure operational safety and enhance production efficiency. Traditional fault feature extraction methods exhibit inherent limitations in composite fault scenarios, particularly when processing non-stationary signals and multi-source fault signals. While intelligent algorithms based on vibration signals have been widely adopted in mechanical fault diagnosis, deep learning methods have gained prominence due to their superior capacities in large-scale data processing, automated feature extraction, and high-dimensional data modeling. This study provides a comprehensive review of vibration signal-based composite fault feature extraction techniques for gearboxes, comparing traditional signal processing methods with modern deep learning approaches. First, classical techniques—including wavelet transform, empirical mode decomposition, symplectic geometric mode decomposition, and blind source separation-are introduced, with an analysis of their respective strengths and limitations in handling nonlinear and non-stationary signals. Subsequently, the application of deep learning models—including convolutional neural networks, deep belief networks, and capsule networks in composite fault feature extraction is examined, comparing their theoretical foundations, performance, and applicable scenarios. Findings indicate that deep learning methods excel in autonomously extracting intricate fault features and demonstrate superior robustness against multi-source interference and noisy environments. However, their effectiveness relies on extensive annotated datasets and substantial computational resources. Future research directions are proposed, including multi-modal data fusion, adaptive learning paradigms, few-shot learning frameworks, enhanced model interpretability, and real-time online monitoring systems. These advancements are expected to drive the evolution and practical implementation of intelligent fault diagnosis technologies.