Residual-Based Hybrid Deep Learning Model for Anomaly Detection in Wireless Sensor Networks

Wireless Sensor Networks (WSNs) are increasingly deployed in environmental monitoring, smart buildings, and industrial IoT applications, where continuous and accurate data collection is essential. However, sensor readings are often affected by anomalies due to environmental noise, hardware malfunctions, or communication issues. Traditional supervised anomaly detection techniques are limited by the lack of labeled anomaly data and the computational constraints of edge devices. This study proposes a lightweight, fully unsupervised hybrid framework that combines Long Short-Term Memory (LSTM) neural networks and the Isolation Forest algorithm for anomaly detection in WSN data. The LSTM model is trained to learn the temporal behavior of temperature readings from individual sensor nodes. The absolute residuals between predicted and actual values are then analyzed using Isolation Forest to identify anomalous patterns. The proposed method is evaluated on the Intel Berkeley Research Lab dataset, which contains real-world WSN temperature readings with irregular sampling intervals. Using synthetic ground truth, the method achieved a detection accuracy of 96.08% on a sparse node with only 61 valid samples, demonstrating its effectiveness in data-limited scenarios. This model requires no labeled training data and is suitable for real-time deployment in resource-constrained environments. The results indicate that the proposed framework offers a practical and generalizable solution for anomaly detection in wireless sensor networks.