With the widespread application of multivariate time-series data in sensor networks, accurately detecting abnormal behavior in the network has become an important research topic. We propose a model that incorporates a Transformer Variant, Graph Neural Networks (GNN)-based feature propagation, and Data Mix techniques (TGM). This model employs a frequency filtering approach with beyond self-attention to capture temporal and frequency features in time-series data while integrating GNN to model the spatial relationships between sensors. By aggregating spatial features and time-frequency features through graph learning, the model’s ability to recognize anomalous patterns is enhanced. To further enhance the robustness of the model, this paper also introduces data augmentation techniques, generating diversified augmented signals through noise injection and signal mixing. It employs a beta distribution-based mixing coefficient and Gaussian noise to enhance the robustness of the signals, helping the model better adapt to uncertainty and noise. In addition, the paper introduces an attention mechanism that dynamically generates the graph structure through the graph attention matrix, thereby improving the model’s ability to capture the complex relationships between sensors. The experimental results show the proposed method can effectively handle complex anomalies in multivariate time series, demonstrating strong robustness and accuracy.