DANA: decentralized agentic neighbor aware anomaly detection in wireless sensor networks
Year: 2026
Type: conference-paper
Abstract: Wireless Sensor Networks (WSNs) are widely used in various monitoring applications. However, the data collected by sensor nodes may be affected by environmental and physical factors, resulting in expected data variations, abnormal changes or sensor faults. This issue is particularly more critical in scheduled based transmission techniques, where a subset of nodes is selected for transmission. In such cases, the transmitted data may not include relevant changes observed by non selected nodes. In such scenarios, a node may observe either a local change, which is specific to its own measurements, or a global change, which affects a group of neighboring nodes. Therefore, enabling each sensor node to locally distinguish between these change types is essential to adapt its transmission decision, avoid missing important information, and prevent the transmission of redundant data. Missing important information will reduce the monitoring precision, while transmitting redundant data increases energy consumption and packet overhead in the network. Consequently, this work focuses on a decentralized agentic based anomaly detection approach that allows each sensor node to reason and update its transmission decision based on the information it perceives. After detecting a local variation in its measurements, a sensor node aims to identify the type of change by relying on overheard data from neighboring nodes. Accordingly, the node can decide whether to transmit its data, suppress its transmission, or wait for additional neighboring information before making a final decision. This process leads to a dynamic sliding window approach, where the observation window is continuously updated according to the perceived neighboring information. Preliminary results show that the proposed approach achieves a precision of 0.89 in distinguishing local from global changes, compared with 0.88 for the fixed sliding window scenario and 0.24 for direct detection, while reducing the average decision delay.
Source: HAL
Institutions: Antonine University, Centre Inria de l'Université de Lille
Language: English
Subfield: Computer Networks and Communications
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7224984110