The article compares several deep learning frameworks via testing under similar conditions in order to apply them in solving real-time problems on various platforms. Traffic signs recognition is an example of such problem. The study compares speed and accuracy of data output of five popular deep learning frameworks: TensorFlow, Neon, CNTK, MXNet, RyTorch. The most popular training models, ResNet-20, IDSIA, ResNet-32, were selected for training in conditions of limited computing resources due to their highly accurate real-time results. Graphics processing units are commonly used to train convolutional neural networks. The article compiled and analyzed the results of calculations using both central and graphics processors, presented methods of frameworks optimization and the results of their operation both on the central processor and on the graphics one. Thus, it is possible to study the data input influence on the accuracy of subsequently obtained result. The study allows determining and optimizing the most suitable framework for traffic signs recognition according to the criteria of output values accuracy and the time spent on models training.

The article discusses methods for automatic text sentiment analysis. The study analyzes the results obtained via the analysis and software implementation of the methods for determining text emotionality. A method based on weight coefficients was proposed for solving the problems in question, and its efficiency was analyzed. Separately, basic algorithms for a text model were discussed and text preprocessing was carried out. The practice section of the article includes an application of the methods on real data, namely, a sentiment analysis of English-language reviews of various films published at www.imdb.com. All methods applied showed 80 % of accuracy in review classification.

The article considers the problem of designing an automated system for monitoring laborato-ry equipment prior to classes at a higher educational institution. The monitoring must be carried out remotely and promptly inform the technical support services should there be found deviations of the measured parame-ters from the required values specified in the regulatory and technical documentation. In addition to that, it is necessary to monitor the condition of the room itself (air temperature, voltage in the electrical network, no water and heat supply systems leakage). Methods of algorithmization and programming of controllers are used to implement the design. Simulation of laboratory equipment failure is carried out via preliminary stage of E-network modeling of the system. The results of the system design include the functional diagram of the equipment monitoring system and the software for the PLC200 controller developed in the CoDeSys tool environment. The study proposes possible ways for further research on the problem in question.

The polarization characteristics of meteorological formations with an equivalent radius from tens of microns to units of millimeters are studied. When irradiating meteorological formations with linear and circular polarization signals, examples of radar images are obtained. The polarization characteris-tics of natural phenomena associated with precipitation are calculated. The article describes an algorithm for measuring the polarization characteristics of meteorological formations using the method of integrated sens-ing. The degree of polarization of signals double scattered by haze, clouds and hail is estimated.

The relevance of studying and optimizing the multiphase signals is determined by their wide use in radiolocation and communications, with quadratic-phase signals being the most common ones. The aim of the article is to study a hardware highly stable method for generating multiphase signals with phase sampling. The article discusses phase shifters based on the elements of digital technology. The results of the study can be used in radiolocation and communications.

The object of the study is analytical and topological methods of calculation of highly struc-tured schemes with cross-couplings. The aim of the study is to substantiate the method of control modeling for assessment of correctness of transfer function of a highly structured system. The study is based on the branch of nonmetric mathematics that is focused on the topological approach to studying highly structured systems. A system of ten components covered with cross-couplings is analyzed. A functional scheme of two branches, the initial structure of a system and transfer function found with the Mason’s rule, is simulated. A unit function is applied to both branches as an input signal. Correct finding of a transfer function is per-formed by using such criterion as equality of output signals graphs of a model’s two branches. The method proposed increases the level of validity of studying highly structured systems.

The article discusses an option of introducing a renewable energy sources (solar power station) into the power supply system of a general education institution located in Surgut, Khanty-Mansi Autonomous Okrug – Ugra. This resulting in fulfillment of requirements of the Federal Law “On Energy Saving, Improving Energy Efficiency, and Amending Certain Legislative Acts of the Russian Federation” No. 261-FZ of 23.11.2009.

The article discusses a problem of constructing a data-driven piecewise linear regression model (also known as Leontief production function, zero elasticity of substitution production function, and fixed proportions production function) with interval uncertainty for the dependent variable. A brief review of application of traditional forms of such models constructed according to the classical point data is given for assessing air quality, analyzing public health’s relation to the agricultural activity, optimizing processes of antibodies’ fragments purification, studying airport capacity, and solving other problems. A sum of approximation errors mode is taken as a loss function. The formulated problem is reduced to the partially Boolean programming problem of acceptable dimension. There should not emerge any calculating difficul-ties when solving the problem due to the existing large amount of acceptable effective software tools. The results of the study can be applied in research using methods of mathematical simulation of complicated technical and socially economic objects with interval uncertainty in the initial data caused by failures in the operation of measuring devices, errors in the activities of statistical services and other reasons.

The article describes a method for searching convolution kernels using the MNIST dataset. The algorithm is studied in detail. The possibility to apply the algorithm is substantiated. The obtained results are analyzed. The algorithm developed will make it possible to substitute discrete kernels, which are selected once for each applied field and do not require training of convolutional layers in neural networks, for trainable convolution kernels consisting of numbers with floating point.

Probabilistic models based on Bayesian networks are used as a common mechanism to describe processes occurring in uncertainty. The optimization of solving factorization problems, distributing evidence, and calculating the complete joint distribution of each vertices of the Bayesian network graph repre-sents a current direction associated with optimizing the calculation of dynamic Bayesian networks’ variational features. The study considers the possibility of presenting Bayesian networks as hypergraphs resulting from the need to develop optimal learning algorithms of Bayesian networks and determine the main approaches to implementation of the network’s survey. Specifics of variational inference in forming transition models applied to certain adjacent time samplings are studied, with semantics of dynamic Bayesian networks considered. Algorithms for discrete and continuous models of dynamic Bayesian networks are presented. The proposed approaches make it possible to optimize the procedure for calculating a priori distributions of a dynamic Bayesian network, simplify its topological structure, as well as optimize the network’s survey when obtaining new evidence and determining the probability distribution for hidden variables according to evidence data.

The article analyzes the problem of searching a set of pairs of modular bases equidistant from the mean base. These pairs provide a minimum of the bivalent effect, and their product covers the required computational range. By finding the bases, binary combinations that cannot be placed in registers for larger bases in a pair can be placed in those for smaller bases in a pair by redistribution of redundancy.