DeepSeaSenseNet: A Hybrid Transformer CNN Framework for Real-Time Anomaly Detection in UIoT Networks
Judy Simon
Abstract
Underwater Internet of Things (UIoT) is emerging as a critical technology for real-time monitoring of marine ecosystems, offshore infrastructure, and environmental phenomena. However, the dynamic and harsh aquatic environment poses significant challenges for anomaly detection due to high latency, low bandwidth, multipath distortion, and frequent node failure. This paper introduces DeepSeaSenseNet, a novel hybrid deep learning framework that integrates Convolutional Neural Networks (CNNs) with Transformer architectures to enable robust and efficient real-time anomaly detection in underwater IoT networks. The CNN module captures local spatiotemporal patterns from sensory data streams, while the Transformer component leverages self-attention to model long-range dependencies and global contextual features. DeepSeaSenseNet is trained on a synthetically augmented underwater dataset, incorporating diverse anomaly types such as equipment malfunctions, salinity fluctuations, and communication disruptions. Experimental evaluations demonstrate that our model significantly outperforms traditional LSTM, GRU, and vanilla CNN approaches in terms of accuracy, precision, recall, and energy efficiency, even under noisy acoustic channel conditions. The proposed architecture also exhibits high generalizability across varying depths and deployment configurations, making it a promising solution for next-generation autonomous underwater surveillance and adaptive event detection systems.