To address the adaptability issue of bearing fault diagnosis models in cross-domain transfer tasks, this study focuses on source domain data selection, preprocessing, and mechanism-driven feature extraction. The Case Western Reserve University (CWRU) Bearing Data Center dataset is selected as the source domain data, covering four states: normal (N), outer race fault (OR), inner race fault (IR), and ball fault (B). Firstly, interference is eliminated through data preprocessing (unifying the sampling rate, applying band-pass filtering, slicing with sliding windows, and standardization). Secondly, based on the bearing fault mechanism, characteristic frequencies (BPFO, BPFI, BSF) are calculated using geometric parameters and rotational speed estimation. Then, multi-dimensional features are extracted from the time domain, frequency domain, envelope spectrum, and time-frequency domain to comprehensively capture fault information. Finally, the effectiveness of the features is verified using a radial basis function kernel support vector machine (RBF-SVM) with class weights, achieving an accuracy rate of 97.14% on the test set and a macro-average F1 score of 0.9705. The results indicate that the extracted features have good generalization and discrimination, laying a foundation for subsequent target domain transfer diagnosis.