Detecting anomalies in time series data is crucial for various applications. However, the unsupervised nature of this task presents significant challenges, particularly in understanding intricate temporal dependencies and identifying patterns across multiple scales without labeled data. Existing methods often struggle to balance sensitivity to subtle anomalies with robustness to normal variations. To tackle these issues, we propose a novel framework, multi-scale Laplace anomaly detection (MSLAD), which explores multi-scale temporal dynamics and association relationships to enhance anomaly detection performance. At the core of MSLAD is the multi-scale Laplace attention mechanism (msLap attention), designed to simultaneously capture dependencies at different temporal resolutions. By leveraging the Laplace kernel, the method effectively emphasizes local temporal structures and regions of significant variability, enabling enhanced sensitivity to diverse and complex anomaly patterns. Additionally, we employ the Wasserstein distance to quantify deviations between normal and abnormal data distributions, offering a robust and meaningful association discrepancy metric. Experimental results demonstrate that the MSLAD model achieves superior anomaly detection performance across diverse datasets, achieving F1-Score on MSL (95. 19%), SMAP (96. 42%) and PSM (97. 89%).