Real-time anomaly detection in IoT streams through spatiotemporal patterns

In the era of rapidly expanding Internet of Things (IoT) ecosystems, ensuring reliable, secure, and real-time anomaly detection remains a pressing challenge due to the high dimensionality, dynamic behavior, and noise-prone nature of sensor data streams. This paper presents a hybrid deep learning framework that integrates Convolutional Neural Networks (CNNs) for spatial feature extraction and attention-enhanced Recurrent Neural Networks (RNNs) for temporal sequence modeling. The architecture employs adaptive thresholding and an incremental retraining strategy to maintain responsiveness under evolving data distributions. Robust preprocessing pipelines, including noise filtering, missing data imputation, and domain-specific rule validation, ensure data quality prior to inference. Extensive evaluation across three real-world IoT datasets demonstrates the proposed model’s superiority over state-of-the-art baselines such as Swin-Transformer, Gated GNN, and ConvLSTM, achieving up to 4.6% higher F1-score, sub-6 ms inference latency, and resilience to noise and network delays with less than 4% performance degradation. Furthermore, integrated interpretability modules—such as attention heatmaps and SHAP-based feature attribution—facilitate root cause analysis for operational trust. These results position the proposed approach as a scalable, efficient, and explainable solution for real-time anomaly detection in large-scale IoT environments.