Abstract: In the era of large models, the training and inference of large models need the support of arithmetic resources, in which the anomaly detection of arithmetic resource data can effectively guarantee the training and inference of large models. As the parameters of the large model increase, the scale of the arithmetic resources used by the large model grows, in which the data of multiple types of metrics reflecting the operating state of computility show more complex temporal changes over time. Existing multivariate time series anomaly detection methods typically use a preset window size to perform sliding slicing on multivariate time series data. However, a unified window that ignores the periodic characteristics of different dimensions may truncate the complete periodic patterns of time series data in some dimensions, hindering the anomaly detection model from learning the normal patterns of multivariate time series data and resulting in poor anomaly detection performance. To address this issue, this study proposes an unsupervised multivariate time series anomaly detection method SELAD based on ensemble learning with multi-window extraction. Specifically, this method first extracts the periodic patterns of each dimension in the multivariate time series data based on the Fourier frequency method, and then performs multi-window extraction to preserve the complete periodic patterns of each dimension. In the process of model training, the huge number of parameters of the large model can solve the problem that the traditional model has a memory bottleneck when the sliding window increases, which leads to the deterioration of the learning effect. Subsequently, by designing a Mixed Expert Models (MoEs), the time series data from multiple partitioned windows are input into an ensemble learning framework that integrates large models and LSTM models for training, in order to learn and identify the normal temporal patterns of each dimension. Finally, anomaly detection is performed based on reconstruction scores. In this study, experiment results on four real-world multivariate time series datasets demonstrate that SELAD improves the F1 score by 17.87% to 90.77% compared to existing methods.