Dual-channel hypergraph networks in the time–frequency domain for learning advanced spatiotemporal dependencies in multivariate time series

Multivariate time series play a crucial role in applications such as human activity recognition and fault diagnosis, where capturing spatiotemporal dependencies is paramount. While recent methods have leveraged graph neural networks (GNNs) or hybrid models that integrate convolutional neural networks (CNNs) and long short-term memory (LSTM) networks to learn these dependencies, two significant challenges remain: First, previous research has neglected the dynamic dependencies and overall correlations among multivariate time series, which are crucial for comprehensive data utilization. Second, when processing both temporal and spatial dimensional information, how to effectively integrate higher-order information while avoiding complexity issues in model design and implementation remains a challenge. To address these challenges, we propose a Dual-channel Temporal–Spatial Hypergraph Network (DTSHN) classification framework based on the time–frequency domain. Specifically, our DTSHN is a time–frequency domain dual-channel interactive learning model that enhances the utilization and diversity of time series data. To capture higher-order relationships among signals, we design a dynamic hypergraph structure learning module and construct temporal and spatial hypergraph structures in the time–frequency domain. Additionally, we propose a temporal–spatial hypergraph network module to learn temporal correlations and spatial dependencies within time series. Furthermore, to address the feature consistency issue in the time–frequency domain under the dual-channel paradigm, we introduce a time–frequency contrastive learning strategy to enable the model to learn effective feature representations across both domains. The experimental results show that our proposed model achieves 96.77%, 94.33% and 97.57% optimal accuracy in HAR, FM and PS datasets, respectively. In addition, we have reduced the FLOPs by 19.8% compared to other optimal models.