Abstract: To address the degradation in diagnostic performance caused by missing sample labels of key components—such as bearings and gearboxes—in mine hoisting machines under complex operating conditions, a fault diagnosis technology for critical hoisting machine components based on multi-scale feature transfer learning was proposed. A Lightweight Fault Diagnosis Model Based on Domain Adversarial and Multi-Scale Time-Frequency Feature Extraction (DAMSF-LFDM) was constructed. A Serpentiform Wavelet Coefficient Matrices (SWCMs) representation was proposed. By combining wavelet packet transform, piecewise aggregate approximation, and serpentiform reorganization, a multi-scale time–frequency feature matrix was constructed to fully capture the internal correlation characteristics of vibration signals across different frequency bands. A Multi-Scale Residual Ghost Convolution Block (MRGCB) was proposed. It employed multiple parallel convolutional layers to effectively extract deep features of the input data at different scales, thereby strengthening the model's ability to capture multi-scale information. To extract personalized fault features from SWCMs and perform adaptive fusion, a Fused Multi-Scale Fault Feature Extraction Module (FMFFEM) was introduced. Feature fusion was carried out via summation, and an adaptive feature-weight allocation mechanism was incorporated to complete the fused extraction of features from different frequency bands. By integrating multi-level maximum mean discrepancy loss with a domain adversarial mechanism, a deep transfer diagnosis network based on the domain adversarial mechanism was established, improving the model's adaptability across operating conditions. Experimental results demonstrated that the DAMSF-LFDM model significantly outperformed the comparative models overall, achieving the highest fault diagnosis accuracy across different transfer tasks. The average cross-condition accuracies on the SEU dataset and the MFS-RDS dataset reached 98.67% and 99.80%, respectively.