In order to improve the performance of anomaly detection for multivariate time series data transmitted in wireless communication networks, a spatio-temporal memory-augmented parallel encoder decoder (STMA-PED) model is proposed. The model adopts a parallel dual-stream coding structure, and achieves the efficient extraction of fine-grained local patterns and long-term dependencies through the local multi-scale feature encoder and the global temporal context encoder. The differentiable neural memory bank is introduced to store the normal pattern prototype, and the reconstruction difference measurement between the encoded representation and the memory prototype is achieved through the attention mechanism. The anomaly scoring mechanism is constructed by integrating the difference index and the distraction index, which significantly improves the accuracy and robustness of anomaly discrimination. The experimental results show that the minimum error of the proposed method is only 0.16% in the calculation of the anomaly score of multivariate time series, and the detection results are consistent with the real annotation in the noise environment of 10 and 60 dB, which verifies its excellent performance in the complex wireless communication environment.