D. M. Klionskiy., A. K. Bolshev
The present paper is concerned with the issue of fault detection in the behavior of complex dynamic objects and systems. The authors offer a new approach to discovering these faults with the help of a one-class classification and neural networks. Furthermore, the neural network structure for the problems mentioned is suggested. The approach developed by the team of authors is tested in order to confirm its effectiveness for detecting computer network invasions and finding faults (abnormal values) in the functioning process of such complex dynamic objects as satellites and flying vehicles. These objects and their systems require an engineer (a controller) to constantly monitor signals obtained from different sensors (temperature, pressure, velocity, acceleration, and vibration sensors) and characterizing the state of an object – normal or abnormal. The number of signals can be very large and each one usually contains a great many samples sometimes numbering several million. Therefore the approach suggested and described in the following is aimed at analyzing objects that have very many signals subjected to the analysis with the aim of finding out whether or not there is a fault in an object. As regards mathematical tools we have used such branches of applied mathematics as neural network development and intellectual analysis of signals on the basis of Data Mining techniques (segmentation methods).