Embedding Intelligence at the Sensor: Naïve Bayes Classifier for Real-Time Fault Diagnosis in Resource-Constrained WSNs

Abstract

The reliable functioning of wireless sensor networks (WSNs), integral to mission-critical Internet of Things (IoT) implementations such as industrial automation and environmental sensing, is significantly undermined by data-centric anomalies. These faults, including drift, offset, stuck-at, gain, and out-of-bounds readings, compromise data integrity, leading to faulty decisions and potential system failures. Although machine learning (ML) offers viable solutions for fault detection and classification, deploying sophisticated models onto sensor nodes with inherent resource limitations presents a persistent challenge. This study introduces an embedded Naïve Bayes (NB) classifier designed for in-situ fault detection and classification, capitalizing on its minimal computational demands and inherent noise resilience. Rigorous evaluation employing diverse real-world datasets with variable fault incidences confirms robust performance across accuracy, precision, recall, and F1-score metrics under strict energy and latency thresholds. Comparative assessment reveals substantially reduced resource utilization (computation, memory) relative to conventional techniques while sustaining high detection efficacy across heterogeneous fault conditions. This methodology effectively reconciles sophisticated analytical capabilities with the processing constraints of edge devices, markedly improving fault tolerance to ensure dependable WSN operation within dynamic IoT settings.