Abstract: Abstract  Satellite telemetry data play a critical role in ground operation and management systems for evaluating the on-orbit performance and health status of spacecraft. Telemetry parameters, generated by sensors and controllers, are inherently multivariate time series characterized by strong correlations, quasi-periodicity, high noise levels, and missing values. Many traditional methods fail to effectively capture temporal dependencies and inter-parameter coupling relationships, making it difficult to identify complex anomalies in multivariate telemetry data and resulting in suboptimal detection performance. To address these limitations, this paper proposes an anomaly detection method based on adversarial autoencoders and spatiotemporal feature fusion. Specifically, multiscale temporal features are first extracted using temporal convolutional networks. Long-term dependency structures among parameters are then constructed by similarity modeling, while short-term interaction structures are built by integrating embedding vectors with sliding-window inputs. A graph neural network is employed to model the inter-parameter relationships. Finally, an improved autoencoder is incorporated into an adversarial training framework to enhance robustness against noise perturbations and missing data. Experimental results on real satellite telemetry data and public datasets demonstrate that the proposed method effectively captures complex spatiotemporal dependencies among telemetry parameters and outperforms existing methods in terms of detection accuracy, significantly improving the accuracy and robustness of multivariate telemetry anomaly detection.