
Keyword: "machine learning"
Vladimir Nikolaevich Kliachkin, Ulyanovsk State Technical University, Doctor of Science in Engineering; graduated from the Mechanical Faculty of the Ulyanovsk Polytechnic Institute; Professor at the Department of Applied Mathematics and Computer Science of Ulyanovsk State Technical University; an author of articles in the field of reliability issues and statistical methods. [email: v_kl@mail.ru]V. Kliachkin,
Dmitrii Anatolevich Zhukov, Ulyanovsk Branch of PJSC ‘Tupolev Design Bureau’, graduated from the Faculty of Information System and Technologies of Ulyanovsk State Technical University; a database engineer at the Ulyanovsk Branch of PJSC ‘Tupolev Design Bureau’; Postgraduate Student of the Department of Applied Mathematics and Computer Science; an author of articles in the field of the statistics methods and machine learning. [email: zh.dimka17@mail.ru]D. Zhukov


Algorithm of Diagnostics of Technical Object Operation Using Aggregated Classifiers
The paper addresses the issue of prediction of technical object’s state of health using the known indicators of its operation. The basic data are the known results of the object state estimation by information about previous service: the technical system is healthy or faulty with predetermined values of specified indicators. Such problem may be solved using the machine learning methods, it reduces to binary classification of states of the object. The study showed that diagnostics quality may be improved by means of the binary classification method including aggregated approach, as well as by means of selection of the volume of the validation set and the method of selection of relevant indicators of object operation. In the example (used for the algorithm testing), Fmeasure value, which is the most reliable diagnostics quality measure for unbalanced classes, has increased by 6% (from 0.86 to 0.91), compared with the Support Vector Machine, and by 2%, compared with the bagging of decision trees which is the best basic method for the example considered (F = 0.89). In some cases, this 2% may be of significance from the perspective of the object operation security. Technical diagnostics, healthy state, operation indicators, machine learning, aggregated approach, validation set, crossvalidation.



Sections: Information systems
Subjects: Information systems. 
Valerii Vladimirovich Kozhevnikov, Scientific Research Technological Institute of Ulyanovsk State University, Candidate of Science in Engineering; graduated from the Pushkin Higher Command School of Radioelectronics; Senior Researcher at the Scientific Research Technological Institute of Ulyanovsk State University; an author of articles in the field of microelectronic system design theory.[email: vvk28061955@mail.ru]V. Kozhevnikov


The Method of Mathematical Modeling of Cognitive Digital Automata
An approach to solving the problem of mathematical modeling of cognitive digital automata (CDA) is proposed. The task of formalizing the concept of the cognitive nature of the CDA mathematical model comes to the fore. The cognitiveness (cognition) of the mathematical model is determined by the possibility of learning and generating solutions that are not provided for in the learning process. A special feature of CDA is that the description of the neural network (NN) structure is used as a structural circuit of the automata, and the logical function "NOTANDOR" is used as the model of the neuron. In the case of the feedbacks formation from the output to the inputs of the neurons, the model of the neuron is a binary trigger with a logical function "NOTANDOR" at the input. As a tool for constructing a mathematical model of CDA, a mathematical apparatus of Petri nets (PNs) is proposed: marked graphs, inhibitory PNs and PNs with programmable logic. The mathematical model is builton the basis of the representation of the CDA in the form of the state equation of the PNs from the class of Murat equations (matrix equations) or a system of linear algebraic equations. The task of formalizing the concept of cognitiveness (cognition) is solved as a result of the logic synthesis (learning) of the initial structural circuit of CDA or the formation of the formula (network algorithm) of CDA. At the same time, the possibility of forming a formula (network algorithm) of CDA depends on the critical mass (quality) of training sets and training algorithms. Hence, the task of generating the minimum set of training sets for a given CDA function or experimentally determined function takes on particular importance. Forecasting or generation of solutions, in turn, is performed on the basis of the mathematical model of CDA obtained in the learning process. intellectual control system, cognitive digital automata, artificial intelligence, neural networks, machine learning, cognition, Petri nets, equation of states, mathematical modeling, synthesis, generation, analysis, logic.



Sections: Mathematical modeling
Subjects: Mathematical modeling, Artificial intelligence. 
Aleksandr Viacheslavovich Mikheev, Ulyanovsk Regional Center of New Information Technologies of Ulyanovsk State Technical University, graduated from Ulyanovsk State Technical University; Software Engineer at the Ulyanovsk Regional Center of New Information Technologies of Ulyanovsk State Technical University; his research interests include machine learning. [email: a.miheev@simcase.ru]A. Mikheev,
Kirill Valerevich Sviatov, Ulyanovsk State Technical University, Candidate of Science in Engineering; graduated from Ulyanovsk State Technical University; Dean of the Faculty of Information Systems and Technologies at UlSTU; an author of 40 articles, 3 monographs, 1 textbook, and 4 State Registration Certificate of the Computer Program; his research interests are in the field of machine learning and robotics. [email: k.svyatov@ulstu.ru]K. Sviatov,
Daniil Pavlovich Kanin, Ulyanovsk State Technical University, a student of the Faculty of Information Systems and Technologies of the Computer Science Department at Ulyanovsk State Technical University, a winner of robotics competitions; his research interests are in the field of machine learning. [email: dankan@mail.ru]D. Kanin,
Sergei Vladimirovich Sukhov, Ulyanovsk Branch of the Kotel’nikov Institute of Radioengineering and Electronics of the Russian Academy of Sciences, Candidate at Science in Physics and Mathematics; graduated from the Ulyanovsk Branch of Lomonosov Moscow State University; Senior Researcher at the Ulyanovsk Branch of the Kotel’nikov Institute of Radioengineering and Electronics of the Russian Academy of Sciences; an author of one monograph, 70 articles, and 2 patents for inventions; his research interests are in the field of optics, computational neurobiology, machine learning [email: ssukhov@knights.ucf.edu]S. Sukhov,
Vadim Georgievich Tronin, Ulyanovsk State Technical University, Candidate of Science in Engineering; graduated from Ulyanovsk State Technical University; Associated Professor at the Department of Information Systems of UlSTU; an author of 50 articles, 1 textbook; his research interests are in the field of project management and theory of inventive problemsolving. [email: v.tronin@ulstu.ru]V. Tronin


The Scene Segmentation in the Tasks for Selfdriving Vehicle Navigation By Using Neural Network Models With Attention
The article deals with the designing process of software module for the road signs recognition. This module is designated for the use in the automated control system of selfdriving vehicles and is being developed at Ulyanovsk State Technical University.The creation of the training dataset sufficient to train neural network models is one of the major tasks to be solved when creating computer vision systems for selfdriving vehicles. In this case, the preparation of a large sample in the semantic scene segmentation task may require considerable efforts for “manual” labeling. Authors describe a convolutional network model with a soft attention mechanism. This network is trained in the classification task with the possibility of extracting an attention mask from the internal network state, which can be used for the semantic image segmentation. This approach allows to reduce datalabeling costs significantly. Artificial intelligence, neural networks, machine learning, computer vision, attention networks.



Sections: Information systems
Subjects: Information systems, Artificial intelligence. 
Aleksandr Viacheslavovich Mikheev, Ulyanovsk State Technical University, graduated from Ulyanovsk State Technical University; Software Engineer of Ulyanovsk Regional Center for New Information Technologies at Ulyanovsk State Technical University; his research interests include machine learning. [email: a.miheev@simcase.ru]A. Mikheev,
Kirill Valerevich Sviatov, Ulyanovsk State Technical University, Candidate of Science in Engineering; graduated from Ulyanovsk State Technical University; Dean of the Faculty of Information Systems and Technologies of Ulyanovsk State Technical University;an author of 40 articles, three monographs, one textbook, and four certificates of registration of computer software programs; his research interests include machine learning and robotics. [email: k.svyatov@ulstu.ru]K. Sviatov,
Sergei Vladimirovich Sukhov, Ulyanovsk Branch of the Kotel’nikov Institute of Radioengineering and Electronics of the Russian Academy of Sciences, Candidate of Science in Physics and Mathematics; graduated from the Ulyanovsk Branch of Lomonosov Moscow State University; Senior Staff Scientist at the Ulyanovsk Branch of the Kotel’nikov Institute of Radioengineering and Electronics of the Russian Academy of Sciences; an author of one monograph, 70 articles, two patents of inventions in the field of optics, computational neurobiology, machine learning. [email: ssukhov@knights.ucf.edu]S. Sukhov,
Daniil Pavlovich Kanin, Ulyanovsk State Technical University, Student of the Faculty of Information Systems and Technologies of Ulyanovsk State Technical University; Winner of robotics competitions; his research interests include machine learning. [email: dankan@mail.ru]D. Kanin,
Iakov Andreevich Akimov, Ulyanovsk State Technical University, Student of the Faculty of Information Systems and Technologies of Ulyanovsk State Technical University; Winner of robotics competitions; his research interests include machine learning. [email: yasha.akimov.73@gmail.com]I. Akimov,
Pavel Mikhailovich Volkov, Ulyanovsk State Technical University, Student of the Faculty of Information Systems and Technologies of Ulyanovsk State Technical University; Winner of robotics competitions; his research interests include machine learning. [email: p.vollkkoff@yandex.ru]P. Volkov


Changing the Lighting in the Scenes By Using Neural Networks Cascades
The article describes the process of creating the artificial illumination of surfaces taking into account shadows on convex and concave surfaces using machine learning algorithms and methods of working with threedimensional objects.Despite the variety of applications for photo processing (LightRoom, Photoshop, Snapchat, Instagram etc.), there is a problem associated with the correct modeling of lighting on surfaces, due to the peculiarity of the photo format. First, the photo is a twodimensional object and contains only indirect information about threedimensional objects depicted on it. Secondly, the photoimage captures a certain state of objects with a certain set of properties, which are not enough for modeling shadows and lighting. The missing information is difficult to recreate by classical algorithms or heuristics because of the large number of parameters for recognition of objects in the photo and their possible combinations. However, machine learning algorithms are able to approximate very complex functional relationships based on training samples. This article describes the solution to the problem of modeling artificial lighting and shadows in photography using several neural networks trained on synthetic data. Artificial intelligence, neural networks, machine learning, computer vision.



Sections: Artificial intelligence
Subjects: Artificial intelligence. 
Kirill Valerevich Sviatov, Ulyanovsk State Technical University, Candidate of Science in Engineering; graduated from Ulyanovsk State Technical University; Dean of the Faculty of Information Systems and Technologies at Ulyanovsk State Technical University; an author of articles, monographs, textbooks, and certifications of software registration in the field of machine learning and robotic engineering. [email: k.svyatov@ulstu.ru]K. Sviatov,
Aleksandr Viacheslavovich Mikheev, Ulyanovsk Regional Center of New Information Technologies at UlSTU, graduated from Ulyanovsk State Technical University; software engineer of the Ulyanovsk Regional Center of New Information Technologies at UlSTU; interested in the field of machine learning. [email: a.miheev@simcase.ru]A. Mikheev,
Maksim Andreevich Shliamov, Ulyanovsk State Technical University, Student of the Faculty of Information Systems and Technologies at Ulyanovsk State Technical University; interested in the field of machine learning. [email: m.shlyamov@simcase.ru]M. Shliamov


Software Package for Visual Monitoring of the Cargo Transportation
The article considers the process of designing and implementing a software package to check the completeness of trucks loading when passing checkpoints using technical vision algorithms and machine learning methods. Load completeness checking is used in the governmental supervision of the Russian roads operation as well as checking the cargo transportation volume by private companies when executing orders under the agreements for cargo transportation services. The article deals with the problem of supervising the volume of transported cargo in opentop trucks with bulk nonmetallic materials (sand, gravel, coal). Currently, the problem of roadfreight transport weightchecking is solved in most cases by the use of weighing units. Also, there are solutions when ultrasonic distance sensors are used that generate a signal corresponding to the shape of the vehicle when driving under several sensors. However, these solutions are quite expensive to install, they do not consider the difference between signals when various vehicle types or models are used on one route. In addition, the weight does not always reflect the completeness of the load body because some materials weigh much more than payload at low volume. And under different weather conditions (e.g. high humidity), the mass of certain goods can vary significantly within 15 percent. At the same time, the transportation efficiency is estimated based on the minimum possible number of passes for the export of cargo of the given volume. The vision systems allow to solve the problem on checking the transported bulk cargo volume. The article focuses on the development description of such a system. Computeraided design, artificial intelligence, neural networks, machine learning, computer vision.



Sections: Information systems
Subjects: Information systems, Artificial intelligence. 
Dmitrii Anatolevich Zhukov, Ulyanovsk State Technical University, graduated from the Faculty of Information System and Technologies of Ulyanovsk State Technical University; Postgraduate Student of the Department of Applied Mathematics and Computer Science; an author of proceedings in the field of the statistical methods and machine learning. [email: zh.dimka17@mail.ru]D. Zhukov,
Vladimir Nikolaevich Kliachkin, Ulyanovsk State Technical University, Doctor of Engineering; graduated from the Mechanical Faculty of Ulyanovsk Polytechnic Institute; Professor at the Department of Applied Mathematics and Informatics of Ulyanovsk State Technical University; an author of scientific papers in the field of reliability issues and statistical methods. [email: v_kl@mail.ru]V. Kliachkin


The Effect of the Control Sample Volume on the Quality of Diagnostics of the Technical Object State
He problem of predicting the serviceability of a technical object in terms of its performance is considered. The source data are the known results of the evaluation of the state of an object based on the results of the previous operation: If the specified values of controlled indicators technical system intact or defective. Such a problem can be solved by methods of machine learning, it reduces to a binary classification of the states of the object. The quality of diagnostics can significantly depend on many factors: the method of training, the correct allocation of factors characterizing the operation of the object, the volume of the sample, and others. The work studies the effect of the control sample volume on the quality of diagnosis, estimated by the number of mispredicted states using the crossvalidation method. The tests were carried out in the Matlab package, ten different training methods were used: logistic regression, support vector method, decision tree bugging, and others. It is shown that the correct choice of the proportion of the control sample can improve the diagnostic quality to 57%. Technical diagnostics, serviceability of the indicator of functioning, machine learning, control sample, crossvalidation.



Sections: Mathematical modeling
Subjects: Mathematical modeling. 
Venera Arifzianovna Alekseeva, Ulyanovsk State Technical University, Candidate of Engineering; graduated from the Faculty of Economics and Mathematics of Ulyanovsk State Technical University; Associate Professor at the Department of Applied Mathematics and Informatics at UlSTU; an author of scientific papers in the field of statistical methods. [email: v.a.alekseeva@bk.ru]V. Alekseeva


The Use of Machine Learning Methods for Binary Classification
The article deals with the problem of objects binary classification. To solve this problem, the use of machine learning methods should be provided. Machine learning is a subsection of artificial intelligence. It is a mathematical discipline using subsections of mathematical statistics, numerical optimization methods, probability theory, and discrete analysis. The goal of machine learning is a partial or complete automation of solutions of complex professional tasks in different fields of human activity such as speech recognition, image recognition, medical diagnostics, diagnostics of technical facilities etc. The use of the following methods for binary classification: decision trees, neural networks, nearestneighbor method, discriminant analysis, Bayesian classifier, a support vector machine, logistic regression, decision trees begging, method of the empirical function and fuzzy inference based on the basis of Sugeno model are proposed. Classification efficiency is assessed with the use of a number of characteristics: meansquare error, ROCcurve, AUC index, etc. In order to improve the accuracy of classes objects prediction, a comparative analysis of the effectiveness of these methods at various cutoffs and a combination of models (socalled aggregate classifier) are offered. Binary classification, machine learning, aggregate classifier, cutoff.



Sections: Mathematical modeling
Subjects: Mathematical modeling, Automated control systems. 
