Towards a more realistic approach to the problem of detecting fuel leaks in filling stations: Mixed time windows

Abstract

The early detection of fuel leaks in filling stations is crucial to minimize environmental risks, such as soil and groundwater contamination. There are some commercial products for fuel leakage detection based on statistical analysis of data from station inventory books. Although they solve the tackled problem, they have some important drawbacks, such as their high price, and issues related to the privacy of station data, which must be shared with the company owning the reconciliation technology.In this work, a solution based on Artificial Intelligence is proposed to address this problem. Machine Learning techniques, specifically two-class supervised classifiers, are applied to data extracted from inventory books of real petrol stations. The classification models used in this paper are trained and tested with real data of days without leaks and simulated data of days with leaks. Thus, the more representative of reality these data are, the better the classifiers will work when implemented in a real filling station. In this sense, the most novel contribution of this paper is the way in which the training sets are constructed to achieve a realistic scenario. These sets are composed of time data windows in which the leak can begin on any day within the window, not necessarily on the first day, as the authors had assumed in a previous contribution. Therefore, they are mixed windows containing a variable number of non-leaking and leaking days.In addition, the design of these data sets complies with the requirements of the current European standard UNE-EN 13160–5. This allows the classifiers to work under even more realistic conditions and thus increase the practical applicability of their results. This work demonstrates that by using two-class classifiers it is possible not only to comply with the standard in terms of the maximum allowable ratio of false positives and false negatives, but also to detect the leak in a shorter time than that established in the norm.