Pattern-wise trust assessment of sensor data
One of the most important tasks of a sensor network (SN) is to detect occurrences of interesting events in the monitored environment. However, data measured by SN is often affected by errors. We investigate the problem of assessing trustworthiness (trust) of a sensor value (tested value) in the presence of events and errors. A usual approach is to express the trust as a deviation of the tested value from a reference value (a normal value). State of the art approaches aim at defining the reference value in terms of a context consisting of values of spatially proximate sensors that are correlated with the tested value. However, they trade accuracy for simplicity and use a fixed context consisting of values of a fixed neighborhood (e.g., all values within a circular neighborhood of radius r). Therefore, such a fixed context fails in most practical cases by under or overestimating the trust values. We present the first pattern-wise method (PW) for trust assessment of sensor data that addresses the limitations of the state of the art approaches by departing from the idea of the fixed neighborhood. We consider a variable neighborhood that consists of an arbitrary subset of the spatially proximate sensors. We define the context as a frequent spatial pattern consisting of values of the variable neighborhood that frequently co-occurs with the tested value in the stream of sensor values. We define the trust as a believe (probability) that the tested value is correct given selected features of the context. We compute trust as the output of the logistic regression, where the input variables consist of the following features of the pattern: (I) the relative frequency; (II) the conditional probability of the tested value given the context and (III) the size of the variable neighborhood. Experimental results confirm the superiority of the proposed method over the state of the art method.