Multivariate time series often suffer from missing values during data acquisition, transmission, and storage due to sensor failures, communication anomalies, and environmental disturbances. This not only undermines data integrity, but also adversely affects downstream analysis tasks such as forecasting and classification. Under two common missing scenarios, namely random missing and continuous missing, existing imputation methods still face several challenges, including distortion of input statistics, insufficient adaptation to different temporal ranges, and limited recovery stability. To address these issues, this paper proposes DyMSI, a dynamic multi-scale decomposition-based method for multivariate time series imputation. Specifically, DyMSI constructs a missing-aware input representation, where normalization statistics are estimated only from non-missing positions to alleviate input distribution shift? it further builds a dynamic multi-scale decomposition framework to adaptively fuse information from different temporal scales, thereby enhancing the model’s adaptability to different missing patterns? moreover, an observation-constrained progressive recovery mechanism is designed to reduce the interference of unreliable intermediate results on subsequent recovery, thus improving recovery stability. Experimental results show that, compared with five existing time series imputation models, the proposed method achieves favorable imputation accuracy under both missing scenarios on five public datasets.