A new methodology for identifying unreliable sensors in data fusion

Abstract

Sensor fusion is a fundamental research topic that has received significant attention in the literature. An important body of research has focused on assessing the reliability of a sensor or more generally an "information source" by comparing the readings with the ground truth in an online or offline manner.The Weighted Majority Voting algorithm [25], a well-known online learning algorithm, is a typical example of a class of approaches that assess the reliability of a sensor by comparing its readings to the ground truth in a online manner. Unlike the latter stream of research, in this article, we tackle the problem of identifying unreliable sensors without the knowledge of the ground truthwhich is a novel research direction in its own right. We advocate that comparing the readings of a sensor to the rest of the sensors gives an invaluable information about its reliability. In this article, we present a solution to the problem based on the theory of S-Model Learning Automata (LA) [17]. Interestingly, the feedback to the S-Model environment LA is defined in an intuitive manner, namely, it is proportional to the number of sensors adhering to the chosen action. Our solution does not impose any constraint on the parity of the number of sensors and thus is general and can handle any arbitrary number of sensors.

Game-Theoretic Learning for Sensor Reliability Evaluation Without Knowledge of the Ground Truth

“…By virtue of analogy with the sensor fusion problem, the approach in [14] foresees two groups of sensors, one group of reliable sensors with the interest in reporting the ground truth and another group of unreliable sensors which has interest in misreporting the truth. In a subsequent work, Yazidi and Herrera-Viedma [20] propose an alternative solution that does not involve the majority voting concept as a way for deducing the reliability of the sensors. Instead of applying a majority-based update such as in [21], Yazidi and Herrera-Viedma [20] propose rather to use a reinforcement learning with continuous feedback as opposed to the binary feedback methodology proposed in [14].…”

“…In a subsequent work, Yazidi and Herrera-Viedma [20] propose an alternative solution that does not involve the majority voting concept as a way for deducing the reliability of the sensors. Instead of applying a majority-based update such as in [21], Yazidi and Herrera-Viedma [20] propose rather to use a reinforcement learning with continuous feedback as opposed to the binary feedback methodology proposed in [14].…”

“…However, the premises of aforementioned two main works [14], [20] for identifying unreliable sensor is a condition according to which the truth prevails over lies expressed using the condition (N R -l)P R +N u p u > (N R +N u )/2 where N R , N u , P R and P u are the number of reliable sensors, number of unreliable sensors, the probability that a reliable sensor reports the truth, and the probability that an unreliable sensor reports the truth respectively. Please note that in the particular case where (P R ,PU) = (1, 0) meaning that a reliable sensors always reports the truth while an unreliable sensor always misreports the truth, the condition reduces to a simple majority condition where the number of reliable sensors constitutes the majority of the sensors.…”

“…Please note that in the particular case where (P R ,PU) = (1, 0) meaning that a reliable sensors always reports the truth while an unreliable sensor always misreports the truth, the condition reduces to a simple majority condition where the number of reliable sensors constitutes the majority of the sensors. It is worth mentioning that the main advantage of the work in [20] compared to the original work [14] is that the former is more general and does not require that the total number of sensors N R + N u is an even number, and therefore we could say that the advantage is relying on more mild condition.…”

“…In this paper, we present a more general solution to the problem of identifying unreliable sensors that does not invoke the condition (N R -l)p R +Nup u > (N R +Nu )/2 commonly used in the original solutions [20], [14]. To solve the problem under those general settings, we resort to the field of game theo1y in order to gradually learn the identity of the sensors in a decentralized manner.…”