In the anomaly detection tasks for multivariate time series， the correlations between variables are complex， and such correlation is difficult to be learned by traditional anomaly detection methods clearly. In addition， most models only consider the correlation between variables， with learning time dependencies insufficiently. Therefore， a time series anomaly detection method based on High-order Feature Aggregation （HFA） was proposed. Firstly， a variable relationship diagram was constructed through graph structure learning. Secondly， the traditional Graph ATtention network （GAT） was enhanced by taking full account of higher-order neighbor node correlations， thereby modeling complex inter-variable relationships more accurately. Finally， temporal dependencies of the series were captured fully through the integration of one-dimensional convolutions with self-attention mechanism. Experimental results on four public datasets demonstrate that compared with the suboptimal baseline model Anomaly Transformer， HFA has the F1 score increased by 1.34% on average； compared with the current mainstream baseline method TopoGDN （Topology Graph Deviation Network）， HFA has the F1 score increased by 9.05% on average. The results of ablation experiments further verify the effectiveness of each module in the model.

Time series data are widely used in fields such as power load forecasting and meteorological analysis. Extracting high-quality representations of time series is crucial for downstream prediction tasks. However， the performance of the existing methods is limited by high-frequency noise interference， difficulty in modeling long-term dependencies， and the scarcity of labels. Therefore， a time series representation method based on Spectral Filtering and Hierarchical Dilation （SFHD） was proposed. Firstly， a Spectral Filtering Block （SFB） was designed to extract multi-scale features through global and local filters， and an adaptive spectral filtering mechanism was used in the frequency domain， so as to weaken the influence of high-frequency noise. Then， a Hierarchical Dilation Block （HDB） was constructed to use exponentially dilated convolutions to enlarge the receptive field progressively， thereby enhancing the ability to capture long-term dependencies. Finally， a change-aware self-supervised pretraining strategy was proposed to force the model to understand underlying structure of the series by masking the highly-dynamic data blocks， thereby alleviating the insufficiency of labeled data. Experimental results on seven public datasets with different prediction lengths demonstrate that， compared with the suboptimal model iTransformer （inverted Transformer）， the average Mean Square Error （MSE） of SFHD decreases by 9.47%， and the average Mean Absolute Error （MAE） decreases by 5.36%. It can be seen that SFHD provides stronger representation capabilities and leads to improved performance on downstream time series prediction tasks.