In the field of AIOps (Artificial Intelligence for IT Operations), unsupervised reconstruction-based anomaly detection models have long been plagued by the identical shortcut problem, which tends to directly replicate the input data instead of learning the intrinsic features for reconstruction. Current solutions to the identical shortcut problem are ill-suited for AIOps due to the unique characteristics of AIOps scenarios, such as complex temporal dependencies, strong correlations among multiple metrics, and non-standard data distributions. To address this challenge, we develop the MUAD model based on the spatio-temporal masking mechanism. By masking time-series data and multi-dimensional metric data, it delves deep into the temporal dependencies and potential correlations among different metrics, enabling more accurate capture of data features. We test MUAD on three public datasets, achieving F1-scores of 0.85, 0.9, and 0.89, respectively. Compared with traditional baseline models, MUAD demonstrates a performance improvement ranging from 38% to 69%.