To accurately evaluate the compaction quality of embankment soil materials, a comprehensive assessment method based on multi-source information fusion of roller acceleration data in both the time and frequency domains is proposed. Quantitative relationship models between compaction degree and different indicators, including compaction energy per unit volume, frequency-domain indicator CV, and time-frequency-domain indicator CWT-THD, were established. Utilizing the Dempster-Shafer (D-S) evidence theory as the information fusion framework, the basic probability assignment values of the three indicators and conflicting evidence were optimized through integration with construction practice and Murphy’s rule. The case validation results demonstrate that the proposed method can improve the accuracy of compaction quality evaluation, avoid the limitations of single-indicator assessments, and enable the analysis of spatial uniformity of compaction quality, thereby achieving rapid and accurate evaluation of soil compaction quality.