SBERT-HCube: A hypercube-based multimodal graph transformer for semantic behavior analysis in IoT networks

Internet of Things (IoT) networks present significant challenges for behavior analysis, continuous monitoring, and semantic anomaly detection due to the massive volume of multi-source data, diverse modalities, and high temporal variability. Key issues such as aligning sensor and numerical inputs, interpreting textual messages, and detecting gradual behavioral drifts highlight the lack of a comprehensive and interpretable multimodal solution. In this study, we propose SBERT-HCube, a hybrid deep learning framework that integrates SBERT-based semantic embeddings, a multidimensional hypercube representation space, and dual-stream transformer modules graph-based and temporal to enable adaptive modeling of IoT node behavior. The framework processes numeric, sensory, and textual data in parallel and combines them through a multimodal attention fusion module to construct dynamic behavioral graphs. These graphs are then used to extract joint embeddings for semantic clustering, anomaly detection, and drift prediction. Experimental evaluations conducted on four benchmark datasets SWaT, WADI, SMAP, and MSL demonstrate that SBERT-HCube consistently outperforms state-of-the-art methods such as DDGCT, M-STGAT, and Stack VAE-G. For datasets without textual information (SMAP and MSL), the text branch was intentionally disabled, and the framework operated using the available numeric and sensor modalities only. It is important to note that the full text-numeric-sensor pipeline is applied only to SWaT and WADI, while SMAP and MSL are evaluated using numeric/sensor data only, as no concurrent textual logs are provided in these datasets. Specifically, the model achieved an F1-score of 0.9652 on SWaT, 0.9088 on WADI, 0.9587 on SMAP, and 0.9561 on MSL, along with correspondingly high precision and recall values. These results validate the model's effectiveness in large-scale, noisy, and temporally dynamic environments, highlighting its potential for scalable, accurate, and interpretable behavior modeling in real-world IoT networks.