Using neural networks for detection of anomalous traffic in automation networks

Opening of local communication means of technological devices towards networks available to the public, supervision of devices, and remote technological device administration are the characteristics of modern automation. As a result of this process, the intrusion of unwanted elements from the Internet into control networks occurs. Therefore, in communication and control networks we have to build in active means to insure access to individual technological process components. This contribution is focused on the insurance of control system data communication via neural network technologies in connection with classical methods used in expert systems. The solution proposed defines a way of data element identification in transfer networks, solves the transformation of their parameters for neural network input, and defines the type and architecture of a suitable neural network. This is supported by experiments with various architecture types and neural network activation functions and followed by subsequent real environment tests. A functional system proposal with possible practical application is the result. The determination of parameters for unambiguous data packet identification, proposal and transformation of the data for neural network input are the most significant outcomes of the contribution. The final experimentally verified proposal itself, as well as subsequent neural network implementation, represent a suitable way for utilization of such a network type in the field. The integration scheme proposal of the dedicated system into real production structure and the execution of the conversion of data elements representing the data packet parameters to be used as an input into the proposed neural 6 network are the main contributions of related research. In addition, software able to carry out conversions for a random file size with continuous communication in the data network recorded was created. Determining data transfer parameters capable of specifying unambiguously the data packets' parameters, and subsequently the neural network proposal, able to detect and control the communication network operation according to parameters selected by us, are essential parts of this contribution as well.