Multi-dimensional time series data describing the same scene or object often exhibit evolving inter-dimensional correlations and intra-dimensional temporal dependencies. Leveraging these dynamic patterns can improve the accuracy of data cleaning. However, existing methods often fail to capture and exploit them, particularly in closely related and time-dependent multi-dimensional time series, leading to suboptimal detection and repair. To this end, we propose Mender, a novel framework for multi-dimensional time series data cleaning based on dynamic patterns, which employs correlation coefficients to capture inter-dimensional influence, incorporates temporal modeling along each dimension, and dynamically performs effective error detection and repair via a sliding window mechanism. Specifically, Mender consists of two modules: detection and repair. The detection phase compares the correlations between dimensions across different windows, and incorporates the stability of speed change rates along dimensions to identify errors. The repair process models dimensional correlations over time as inter-dimensional influences and contributions to construct a repair candidate set, from which the final repair values are determined via a clustering-based optimization strategy. Experiments on real-world datasets demonstrate that Mender achieves higher error detection accuracy and generates repaired values closer to the ground truth than existing methods.