A metric-correlation-based fault detection approach using clustering analysis in wireless sensor networks

Fault detection plays a crucial role in wireless sensor networks (WSNs). Many fault detection approaches requiring a priori knowledge of network faults have been proposed to distinguish faulty sensors by exploring spatial-temporal correlations among sensor readings. However, many faulty sensors that may not generate anomalous sensor readings, and potential failures with unknown types and symptoms remain undetected. In this paper, we propose a Metric-Correlation-Based Fault Detection (MCFD) approach using clustering analysis. It is motivated by the fact that the system metric correlations of most fault-free sensors usually show strong similarities, whereas different patterns of such correlations indicate potential failures. MCFD explores internal metric correlations inside sensors using correlation value views. An improved Neighbor-based Local Density Clustering Analysis (NLDCA) algorithm based on the Neighbor-based Local Density Factor (NLDF) is applied in spatial domain detection to cluster similar correlation value views together, thus potential faulty sensors with abnormal views not belonging to any cluster can be detected. Simulation results demonstrate that MCFD approach performs well in respects of higher detection accuracy and lower false positive rate even under high node failure ratios and dense distribution conditions.