Learning causal dependencies to detect and diagnose faults in sensor networks

Abstract

Exploiting spatial and temporal relationships in acquired datastreams is a primary ability of Cognitive Fault Detection and Diagnosis Systems (FDDSs) for sensor networks. In fact, this novel generation of FDDSs relies on the ability to correctly characterize the existing relationships among acquired datastreams to provide prompt detections of faults (while reducing false positives) and guarantee an effective isolation/identification of the sensor affected by the fault (once discriminated from a change in the environment or a model bias). The paper suggests a novel framework to automatically learn temporal and spatial relationships existing among streams of data to detect and diagnose faults. The suggested learning framework is based on a theoretically grounded hypothesis test, able to capture the Granger causal dependency existing among datastreams. Experimental results on both synthetic and real data demonstrate the effectiveness of the proposed solution for fault detection

Making Intelligent the Embedded Systems Through Cognitive Outlier and Fault Detection

“…To learn , we apply the Granger-based theoretical framework presented in [5] to the set Z N . There, relationships between datastreams are assessed through the theoretically-grounded Granger causal dependency test.…”

“…In the synthetic dataset, we model an intelligent embedded system whose dependency graph has the following characteristics || = 4 and |E| = 4 as described in [5] and where the data generating process is:…”

“…In the interval [1842; 2861] we injected outliers uniformly, with probability γ i ∈ {0.01, 0.05}, whose signature is equal to (t) = x i (t init ), for t ≥ t init . In the considered solutions the confidence of g of the dependency graph learning procedure was set to 0.01, while for the outlier detection and isolation phases we set Γ = 2, = 0.7,̄D i = 3 and̄I i = 2 ∀i, and the parameter C of the HMM-CDT was set to 4.1 (the estimation procedure of the threshold was conducted as in [5]). We average the results over 100 independent runs.…”