Smart city infrastructure protection: real-time threat detection employing online reservoir computing architecture

Abstract

The most important problems that occur during the extraction of knowledge from data streams are related to the properties characterizing “BigData,” namely high speed of information flow (velocity), variety of used forms, variability of data and diversity of information accuracy diagnosis methods (veracity). The use of online or sequential learning methods offers a specialized solution for solving real-time data processing problems. Data are provided without a clear knowledge of their particular inherent characteristics. Conventional approaches focus on applying heuristic or logical analysis rules. They fail to effectively handle new patterns (produced as a function of time) and to consider the dynamic change rate of their characteristics. In most cases, these methods approximate, by creating general rather than clear imprints of knowledge, which is hidden in the flows. Moreover, their function requires significant computational resources. This paper introduces (to the best of our knowledge, for the first time in the literature) the implementation of a specialized online reservoir computing architecture for smart city infrastructure protection which has low requirements in computing resources; it is efficient and suitable for real-time data flow analysis. More specifically, it describes the development of an echo state network, comprised of analog neurons with sparse random connections at the input levels and at the dynamical reservoir. Its training at the output level is performed with the recursive least square method. A complex data set was selected for the testing of the proposed model, which fully simulates the digital attacks that can be faced by the mechatronic equipment used in smart water supply networks located in the front end of the smart cities.