The System Simulating the State of the Objects and the Process of Their Monitoring with the Help of the Convolutional Neural Network

Abstract

In this paper, the implementation of the system simulating the states of the objects and the process of their monitoring is considered. The tasks for this system include detection and diagnosis of defective states of the studied objects. In order to configure flexibly the system and to reduce the human factor, the use of neural networks is proposed. The model of the complex for the object monitoring simulation is presented. It is introduced in the form of a software application and is tested as a part of a diagnostic system fitted in the railroad track measuring car. This system makes it possible to determine the dimensions of artificial structures on the railway and, most importantly, to find the sites where the standard overall dimensions are reduced due to the deviation of separate structural elements from their designed position. The algorithmic and mathematical support employed in this paper allows determining the geometric (dimensions and position in space), the kinematic (speed and direction of movement) and the dynamic (motion acceleration, unredeemed car acceleration) characteristics of both moving and stationary objects. During its implementation, such system takes the form of a software and hardware complex equipped with movable and fixed photodetectors, depending on the set of tasks to be solved. A characteristic feature of this study is that the methods and the algorithms it proposes for image processing can be adapted and used in building digital twins of transport infrastructure objects including not only the railway tracks but also various technological and service facilities. Due to this approach, the input of the nodal points of infrastructure objects and their elements can be automated into the project design scheme, exponentially reducing the laboriousness of creating primary models. The aggregation of the proposed methods with the capabilities of modern neural networks improves the primary image processing speed on the side of the client application, without fully occupying the data transfer channel between different modules of the information-measuring complex.