A reproducible neuro-fuzzy framework for interpretable anomaly detection in wireless sensor network data streams

Abstract Anomaly detection in wireless sensor networks plays an important role in providing a reliable mechanism for environmental monitoring, healthcare, and industrial automation. Statistical as well as clustering techniques have their own restrictions regarding adaptability whereas machine learning and deep learning algorithms may suffer from high computational costs and low interpretability. To address these limitations, we propose a novel two-phase reproducible anomaly detection framework which includes the combination of Probabilistic Adaptive Neural Gas (PANG) with the Dynamic Evolving Neuro-Fuzzy Inference System (DENFIS). In the first phase, PANG enhances traditional Growing Neural Gas approach via probabilistic expansion, posterior-dependent pruning, and uncertainty-aware error updating to provide prototypes which characterize normal and anomalous behavior of sensors. The second phase uses PANG prototypes, Mahalanobis distance, spatial and attribute correlation, and temporal deviation to build membership functions and evolving rules for classification into the categories of Normal, Suspect and Anomalous using data-driven fuzzy membership function estimation approach with ordered fuzzy partitioning and overlapping Gaussian functions. The proposed algorithm outperforms baselines such as GNG, DBL-GNG, and deep autoencoder models on Intel Berkeley Research Lab and ISSNIP indoor-outdoor WSN datasets achieving 99.0% accuracy on IBRL and 98.8% on ISSNIP, with false alarm rates below 2%.