+7 (495) 987 43 74 ext. 3304
Join us -              
Рус   |   Eng


Rysina (Lobaneva) Ekaterina I.

Postgraduate, Applied Mathematics and Artificial Intelligence Department, National Research University "MPEI"
Moscow, Russia

Algorithms for the formation of images of the states of objects for their analysis by deep neural networks

Algorithms of visualization of numerical data characterizing the state of objects and systems of various nature with the aim of finding hidden patterns in them using convolutional neural networks are presented. The algorithms used methods for obtaining images from numerical data on the basis of the discrete Fourier transform of time series fragments, as well as on the basis of the application of visualization using three-component system diagrams, if such a three-component representation of the system is possible. The software implementation of the proposed algorithms was performed in the Linux environment in the Python 3 language using the Keras open neural network library, which is a superstructure above the TensorFlow machine learning framework. For the learning process of the neural network, a Nvidia graphics processor was used that supports the technology of the CUDA parallel computing software and hardware architecture, which significantly reduced the learning time. The proposed approach is the recognition States of the objects according to their visualized data are based on the recognition of no boundaries or forms of the figures in the images and their textures. Also presented is a program that generates sets of images to implement the process of learning and testing convolutional neural networks in order to pre-tune them and assess the quality of the proposed algorithms.Keywords: Internet, Internet security, parental control applications, user security, information security, Internet threats.

Preliminary assessment of the pragmatic value of information in the classifiсation problem based on deep neural networks

A method is proposed for preliminary assessment of the pragmatic value of information in the problem of classifying the state of an object based on deep recurrent networks of long short-term memory. The purpose of the study is to develop a method for predicting the state of a controlled object while minimizing the number of used prognostic parameters through a preliminary assessment of the pragmatic value of information. This is an especially urgent task under conditions of processing big data, characterized not only by significant volumes of incoming information, but also by information rate and multiformatness. The generation of big data is now happening in almost all areas of activity due to the widespread introduction of the Internet of Things in them. The method is implemented by a two-level scheme for processing input information. At the first level, a Random Forest machine learning algorithm is used, which has significantly fewer adjustable parameters than a recurrent neural network used at the second level for the final and more accurate classification of the state of the controlled object or process. The choice of Random Forest is due to its ability to assess the importance of variables in regression and classification problems. This is used in determining the pragmatic value of the input information at the first level of the data processing scheme. For this purpose, a parameter is selected that reflects the specified value in some sense, and based on the ranking of the input variables by the level of importance, they are selected to form training datasets for the recurrent network. The algorithm of the proposed data processing method with a preliminary assessment of the pragmatic value of information is implemented in a program in the MatLAB language, and it has shown its efficiency in an experiment on model data. Read more...

Algorithm for predicting the parameters of a system for processing waste apatite-nepheline ores

Within the framework of the concept of a circular economy, research in the field of creating technological systems for recycling waste from mining and processing plants occupies one of the key positions. This is connected, on the one hand, with significant volumes of such waste, reaching tens of millions of tons and posing a significant environmental hazard to air and water basins, human health, and, on the other hand, with their rich chemical and mineralogical composition, which makes it possible to call them accumulations of technogenic deposits. In this regard, the task of creating control systems for technological processes of processing such waste and their information support, including support for all stages of the passage of information processes, is urgent. The novelty of the presented research lies in the proposed structure of an intelligent control system for a complex chemical and energy technological system for processing apatite-nepheline ores, as well as in an algorithm for predicting technological parameters, which is part of the information support of the control system under consideration. The algorithm is based on the use of the apparatus of deep recurrent neural networks and Kalman filtering, which is used at the stage of data preprocessing to train the neural network. The paper describes the proposed algorithm for predicting multidimensional time series, adapted to the considered technological process, presents the software executed in the MatLab environment to demonstrate the efficiency of the specified combination of methods for processing technological parameters. In a model experiment, it has been shown that the use of filtering makes it possible to increase the accuracy of the forecast, which is especially noticeable at its large horizons. The practical significance of the research results is the proposed structure of an intelligent control system for the processing of apatite-nepheline ore waste and software for predicting its parameters, which can be used in various decision support systems. Read more...

Determination of informative spectral ranges for the development of a transformer oil control system using deep learning neural networks

Optical spectral methods in the ultraviolet and visible regions can be used to develop transformer oil control technologies based on deep learning neural network models. The aim of the research is to identify informative spectral ranges of luminescent diagnostics for the automation system for monitoring the characteristics and parameters of transformer oil using deep learning neural networks. Measurements of the spectral characteristics of pure and spent transformer oil in the range of 180-700 nm were carried out on a diffraction spectrofluorimeter "Fluorat-02-Panorama". A qualitative and quantitative difference in the excitation spectra has been established: for waste oil, the spectra are shifted to the right and reduced by about four times to the maximum. The excitation maxima are located at wavelengths of 300, 322, 370 nm for pure and 388, 416 and 486 nm for waste oil. The photoluminescence spectra of pure oil at 300 nm excitation are a superposition of at least three curves, the largest of which has a maximum at 382 nm. For excitation of 370 nm, the spectrum is significantly wider and has maxima at wavelengths of 387, 405, 433-439 and 475-479 nm. The photoluminescence spectra of used oil are several times lower and have maxima at 446, 483 and 520-540 nm. The established excitation and luminescence ranges will be used when creating a methodology and installing quality control parameters of transformer oil during its operation. A deep learning neural network model based on the use of a self-organizing Kohonen map was also developed, which made it possible to predict the spectral characteristics of excitation based on the photoluminescence flow of transformer oil and, as a result, to determine the efficiency of the described method in industry through a decision-making system. Read more...

A method for classifying mixing devices using deep neural networks with an expanded receptive field

The paper presents the results of research aimed at developing a method and software tools for identifying the class of a mixing device by its resistance coefficient through experimental data processing. Currently, the main methods for studying mixing devices are finite element methods, as well as procedures of estimating turbulent transfer parameters using laser dopplerometry and chemical methods of sample analysis. These methods require expensive equipment and provide results only for certain types of equipment. This makes it difficult to extend the inferences to a wider class of devices with different designs of mixing impellers. The proposed method involves processing the results of an experiment in which a point light source forming a beam directed vertically upwards is located at the bottom of a container filled with a transparent liquid. A mixing device with variable rotation frequency is placed in the container. When performing experiments in real conditions, small deviations in the size and location of the mixing device lead to difficult-to-predict fluctuations of the funnel surface. Therefore, the image of one marker describes a trajectory that is difficult to predict. It, under certain conditions, can intersect with the trajectories of other markers or be interrupted at the moment when the marker is closed by a stirrer blade passing over it. The resulting image of the markers is associated with a change in the rotational speed of the blade by a rather complex relationship. To identify this dependence, it is proposed to use deep neural networks operating in parallel in two channels. Each channel analyzes the video signal from the surface of the stirred liquid and the time sequence characterizing the change in the speed of rotation of the blades of the device. It is proposed to use neural networks of various architectures in the channels - a convolutional neural network in one channel and a recurrent one in another. The results of the operation of each data processing channel are aggregated according to the majority rule. The computational novelty of the proposed algorithm lies in the expansion of the receptive field for each of the networks due to the mutual conversion of images and time sequences. As a result, each of the networks is trained on a larger amount of data in order to identify hidden regularities. The effectiveness of the method is confirmed by testing it with the use of a software application developed in the MatLab environment. Read more...