Search

Search Results

• OPTIMIZATION OF THE COMPUTATIONAL SCHEME FOR THE INTERPOLATION OF DECADAL METEOROLOGICAL DATA BY INVERSE DISTANCE WEIGHTING WITH PARALLEL PROCESSING OF MULTIPLE TIME SLICES

  О.М. Golozubov , А.V. Kozlovskiy , E.V. Melnik , Y.E. Melnik , А.N. Samoylov

  22-32

  2025-12-30

  Abstract ▼

  The present study is devoted to solving the problem of computational inefficiency in spatial interpolation of large arrays of decadal meteorological data using the inverse distance weighting method. Traditional approaches involving sequential and independent processing of each time slice demonstrate a linear increase in execution time and significant RAM consumption, which becomes a critical barrier to the rapid construction of detailed and geographically linked raster fields in GeoTIFF format. This significantly limits the use of the method in tasks requiring rapid processing of long-term data archives. The purpose of this work is to develop and validate an optimized computational scheme that can radically reduce time costs while maintaining the completeness and accuracy of the results. The key scientific novelty of the proposed approach lies in the fundamental rethinking of the computational process. Instead of repeating identical operations many times, a scheme is proposed based on a single calculation of the full vector of geodetic distances from each grid cell to all weather stations. This most resource-intensive operation is performed only once. Subsequently, the resulting distance vector is applied to all time slices (decades) to calculate the interpolated values, which eliminates the main computational redundancy and ensures a sublinear dependence of processing time on the number of decades. To further improve performance, a parallel processing mechanism is used at the CPU level, implemented by dynamically dividing the raster into independent computing units (batches). The size of the batches is adaptively adjusted taking into account the available RAM, which guarantees the stability and scalability of the solution on systems of various capacities. The testing of the method on real meteorological data for the period 2015-2024 demonstrated a radical reduction in the execution time. In particular, processing ten decade time slices on a standard laptop takes less than 3.5 minutes, and on a server platform it takes about 3 minutes, which represents a multiple acceleration compared to traditional implementations. Thus, the developed solution makes the operational processing of large spatial and temporal meteorological arrays a reality for a wide range of researchers, opening up new opportunities for climate monitoring, agrometeorology and geoinformation analysis without the need for specialized expensive equipment

• METHOD OF MOVING OBJECT POSITIONING WITHOUT USING GLOBAL GEO-REFERENCED DATA

  Е. V. Lishchenko, E.V. Melnik, А. S. Matvienko, А.Y. Budko

  2025-01-09

  Abstract ▼

  The paper considers the problem of determining the current coordinates of moving object in the
  conditions of unstable signal from the global navigation satellite system (GNSS). The relevance of the
  work is due to the fact that in recent years moving object are increasingly used in virtually all sectors of
  industry, agriculture, transportation, solving a variety of tasks of surveillance, reconnaissance, monitoring
  the state of controlled objects, search and rescue operations, cargo delivery and much more. At the same
  time, the success of flight missions largely depends on how accurately and efficiently its onboard navigation
  system works in real time. The existing solutions for creating onboard positioning systems involve the
  use of inertial and GNSS. However, they have the disadvantage of partial or complete absence of data
  from the GNSS (Global Positioning System). This paper describes a method for maintaining a given accuracy
  of moving object spatial positioning under conditions of partial or complete absence of data from the
  object's GSP. This approach is based on a combination of computer vision methods for processing video
  stream frames from the moving object on-board vision system (OVS) in order to ensure positioning accuracy
  under conditions of partial or complete absence of data from satellite navigation systems. Based on
  the advanced method, an algorithm has been developed for automated determination of moving object
  coordinates in the absence of georeferencing data from global positioning systems (GPS). Experiments
  have been carried out, which demonstrated the reduction of time costs for description and matching of key
  points and improvement of the accuracy of image matching. The developed algorithm was used to solve
  the problem of satellite image matching, which is an important step in the moving object positioning problem
  without the use of global geo-referencing data.

• DESIGNING MLP AND CNN NEURAL NETWORK MODULES ON FPGA FOR IMAGE CLASSIFICATION TASKS

  E. V. Melnik , D.Е. Blokh , А.I. Bezmeltsev , V.S. Panishchev , S.N. Poltoratsky

  214-229

  2025-11-10

  Abstract ▼

  Relevance. The development of machine learning methods and neural network architectures, as well as their spread into various industrial sectors, determine the relevance of solving problems related to their hardware implementation. The use of programmable logic integrated circuits in this area will increase data processing speed and the adaptability of the implemented algorithms. However, designing neural network architectures on programmable logic integrated circuits is associated with a number of methodological and technical difficulties, including the optimization of parallel computing, hardware resource management, and ensuring operation under conditions of limited computing resources. The purpose of this work is to analyze and compare two neural network architectures, the multilayer perceptron (MLP) and the convolutional neural network (CNN), in the context of their hardware implementation on programmable logic integrated circuits (PLICs). Particular attention is paid to the trade-off between classification accuracy and the efficient use of limited FPGA hardware resources. Research methods.
  To achieve the goal, two modules were developed and simulated on a Virtex 7 FPGA, a perceptron and a convolutional module. The MNIST dataset, reduced to 20×20 pixels, was used. The implementation included quantizing parameters to a fixed 16:16 format, optimizing hyperparameters, using tabular computations for nonlinear functions, and evaluating FPGA resource usage. Results and discussions.
  MLP achieved 93% accuracy using 11% of logic elements, while CNN achieved 98% accuracy but required significantly more resources. The use of internal buffers to store intermediate data in CNN resulted in exceeding the allowable resources. The forced transition to external memory increased delays and the number of I/O ports. Conclusions. The study showed that the choice of architecture depends on priorities: CNN provides better accuracy but is less resource-efficient. For embedded systems with memory and power consumption constraints, a simplified MLP implementation is preferable. The main problems remain the lack of internal memory and the high resource intensity of operations, which requires further research in the field of hardware optimization and adaptive computation control

• INVESTIGATION OF STRUCTURAL CHARACTERISTICS OF DISTRIBUTED COMPUTING SYSTEMS BASED ON GRAPHS WITH MULTIPLE EDGES OF DIFFERENT TYPES

  E.R. Muntyan , E.V. Melnik

  2021-08-11

  Abstract ▼

  The article considers the issues of fault-tolerant computing systems (CS) development in
  terms of their structure and redundancy. It is necessary to take into account a great amount of
  factors, which have an impact on performance, reliability and fault tolerance, during the distributed
  CS development. For distributed CS such factors contain, among other things, structural characteristics.
  The dependency graphs of processor nodes (PN)nonfailure operating probability of
  distributed CS against system structural characteristics are presented in the article. The application
  of advanced redundancy methods, such as performance redundancy, increases the complexity
  of structure design problem. In the case of performance redundancy, instead of adding redundant
  nodes to the system, it is proposed to use redundant computational resources among the involved
  processor nodes. If a node fails, its tasks are reallocated to a free reserve of nodes, which are able
  to work. To implement this method of system redundancy, the organization of a multi-program
  operation mode is required, when some tasks can be performed simultaneously on each node.
  Need for multi-program operation mode providing leads to increasing the number of system configurations, which have to be analyzed on the design stage and in the case of reconfigurations
  when a failure takes place. To reduce the labour intensity of a configuration analysis an approach
  based on graphs with multiple edges of different types is proposed. The use of models based on
  such graphs makes it possible to represent the structure of CS taking into account a multi-program
  operation mode and at the same time significantly reduce the computational time of basic characteristics
  by means of applying relations in the form of a vector, which allows to integrate some
  relations of different types.

• AN APPROACH TO BUILDING ADAPTIVE OBJECT ACCOUNTING SYSTEMS USING ARTIFICIAL INTELLIGENCE METHODS

  V.I. Voloshchuk, Ali Garyagdiyev, М.А. Kozlovskaya, Y.E. Melnik, А.N. Samoylov

  2024-11-10

  Abstract ▼

  The use of artificial intelligence methods for object accounting is associated with a number of difficulties,
  such as the variability of objects, the influence of shooting conditions, the overlap of objects in
  complex scenes, the need to work with different scales and high accuracy, as well as the presence of noise
  distortions in the data. The paper proposes an approach based on dynamic learning and adaptation to
  input data to organize the setup and operation of adaptive object accounting systems based on artificial
  intelligence methods, which includes several consecutive stages. The first stage is the semantic analysis of
  the user's request, which is based on the use of vector-graph data structure, which provides the allocation
  of semantically important elements of the request, allowing the system to understand the context of the
  task and adapt the strategy of search and classification of objects. Then follows the stage of automatic collection and preprocessing of data from open sources, which provides the expansion of the training
  sample and increases the stability of the model. The next important step is the generation of the training
  sample. This process includes image retrieval based on query semantics, manual validation and data partitioning,
  and initial training of the system for automatic partitioning. The above steps are repeated until
  the desired system performance is achieved. The iterative process of pre-training based on alternation of
  automatic markup and manual correction allows to reduce time expenditures on formation of training
  samples. The advantage of using vector-graph structure is the formation of more accurate semantic representation
  of information. Data augmentation including rotation, reflection, scaling, changing brightness
  and contrast, and adding noise is applied to enhance the generalization ability of the model. The proposed
  approach is designed to improve the efficiency (as the ratio of system operation time to its setup time) of
  object registration systems, ensuring their adaptability to different tasks and survey conditions

• A HARDWARE-ORIENTED METHOD OF ACCELERATED SEARCH BY TEMPLATE BASED ON STRUCTURAL-PROCEDURAL COMPUTING

  Е. А. Titenko, E.I. Vatutin, М.А. Titenko, А.P. Loktionov, E.V. Melnik

  2024-11-10

  Abstract ▼

  The operation of searching for occurrences of a pattern in a text is generally significant in modern
  computing tools for solving problem-searching tasks. Of greatest interest are hardware and software solutions
  that have a homogeneous structure and regular connections between computing blocks. The aim of
  the work is to reduce the time costs for searching for occurrences based on the use of parallel search in
  associative memory and the method of parallelization by iterations. The proposed method uses associative
  memory for parallel search for occurrences and dynamic reconfiguration of the structure of the original
  string from a one-dimensional form to a matrix form. The method is critical to such resources as the number
  of memory access channels, the volume of block memory for creating and parallel operation of an
  array of associative cells. Involvement of all elements in the reconfiguration entails excessive costs of the
  internal block memory for sequential viewing of partial entries by one set of starting positions multiples of
  the sample length (the second symbolic operand). Instead, an approach is proposed to combine in time the
  search for partial entries by two sets of substrings multiples of the sample length, with a simultaneous
  proportional reduction in the elements of the bit slice of the associative memory for each set, which allows
  processing several sample symbols at the current search step. Quantitative estimates of search time are
  determined by the number of comparison and substring writing operations in the overall work cycle, as
  well as the proportions of the time of these operations. It is shown that for samples of more than 10 elements,
  the time gain is approximately 1.8-2 times. This effect is obtained by eliminating the steps of sequential
  shift with transitions between the boundary elements of the strings. The developed method provides
  pipeline processing of a stream of string operands with a combination of viewing at the current
  search step of a non-unit set of characters of the processed string. The search time is re duced by introducing
  a pipeline, the number of stages of which depends on the reduction coefficient of the bit slice size,
  which allows hardware implementation of the structural-procedural approach used in reconfigurable
  computing systems

• APPLICATION OF A HYBRID NEURAL NETWORK AE-LSTM FOR ANOMALIES DETECTION IN CONTAINER SYSTEMS

  I.V. Kotenko, М.V. Melnik

  2024-11-10

  Abstract ▼

  The popularity of container systems attracts the attention of many researchers in the field of information
  technology. Containerization technology allows to reduce the cost of computing resources when
  deploying and supporting complex infrastructure solutions. Ensuring the security of container systems and
  containerization in general, as well as the use of smart attacks based on artificial intelligence by malefactors,
  is a serious problem on the way to the safe and stable operation of container systems. This article
  proposes an approach for detecting not only previously unknown individual anomalous processes, but also
  anomalous process sequences in container systems. The proposed approach and its implementation based
  on the Docker platform are based on tracing system calls, constructing histograms of running processes,
  and using the AE-LSTM neural network. The process of constructing histograms is based on accounting of
  the number of executed system calls for each individual process. This solution provides the ability not only
  to accurately identify any process in the system, but also to effectively detect anomalous process sequences
  with a high degree of accuracy. The generated sequences are used as input data for the neural network.
  After completing the training process, the neural network acquires the ability to detect anomalous sequences
  by comparing a given threshold of reconstruction error with the actual error level of the input
  data vector. When the neural network encounters a new input data vector, it calculates the reconstruction
  error level - the difference between the expected and actual value. If this error exceeds a predetermined threshold, the system signals the presence of an anomaly in the sequence. Experiments show that the proposed
  approach demonstrates high accuracy in detecting anomalous processes with a low level of false
  positive detection results. Such results confirm the effectiveness of the proposed approach. Also, the computational
  costs of training the neural network model are quite low. This allows using less powerful hardware
  without significant performance losses. Such a solution can be trained and implemented in a new
  infrastructure in a fairly short time

• LARGE LANGUAGE MODELS APPLICATION IN ORGANIZATION OF REPLENISHMENT OF THE KNOWLEDGE BASE ON METHODS OF INFORMATION PROCESSING IN SYSTEMS OF APPLIED PHOTOGRAMMETRY

  А.V. Kozlovskiy, E.V. Melnik, А.N. Samoylov

  2024-08-12

  Abstract ▼

  The article deals with the issues related to the automation of the procedure of synthesis of applied
  photogrammetry systems. Such systems serve to measure and account for objects from images and are
  now widely utilized in various fields of activity, such as mapping, archaeology and aerial photography.
  Increasing availability and mobility of imaging devices has also contributed to the widespread application.
  All this has led to active research aimed at developing methods and algorithms for applied photogrammetry
  systems. Manual tracking of new methods and algorithms of photogrammetric information
  processing for a wide range of application areas is quite difficult, which makes the automation of this
  procedure urgent. The solution proposed in the article is based on the use of a knowledge base of information
  processing methods in applied photogrammetry systems, the main elements of which are a fuzzy
  ontology of the subject area and a database, which is logical, since the information about the subject area
  can be structured quite easily. As a basis for the ontology, an existing solution was taken, which was supplemented
  based on the results of analyzing the current state of the subject area. The resulting ontology
  was further used to search and classify information processing methods in applied photogrammetry systems
  and to populate the knowledge base. Due to the intensification of the development of new methods of information processing in the systems of applied photogrammetry, there is a need to modify the ontology
  and to replenish the database, i.e. to replenish the knowledge base. The Internet is an important source of
  information for this purpose. To automate the search for data on information processing methods and
  ontology modification, it is reasonable to use large language models. To automate data mining of information
  processing methods and to populate the knowledge base, it is useful to use large language models
  that simplify several natural language processing tasks, which include clustering and formation of new
  entities for classification. The corresponding method is described in the paper. For the method the results
  of testing its performance are given. As part of problem solving, a comparative analysis of large language
  models has been carried out, resulting in the RoBERTa model.

• ADVANCED PRODUCTION OUTPUT ENGINE FOR IMPLEMENTING PARALLEL COMPUTING

  Е.A. Titenko, I.Е. Chernetskaya, М.А. Titenko, E.V. Melnik, D. А. Trokoz

  2024-05-28

  Abstract ▼

  Relevance. The paper discusses a theoretical approach to organizing parallel computing based on a
  production model of data flow control. The production paradigm of parallel computing has the necessary
  conditions for building new architectures and organizing high-performance parallel computing. We consider
  production (mathematical) systems that control sets of left-hand sides of productions (samples). The
  goal is to increase the efficiency of parallel inference of solutions by reducing unproductive time spent
  searching through possible alternatives in the inference graph space. The research is based on the creation
  of an extended symbolic computation machine for implementing parallel steps. A symbolic computing
  machine is an abstract system that systematizes production output as a sequence of four computational
  and search stages. The inference engine defines the general appearance of a homogeneous computing
  system. The main difference is the decomposition of the base of production rules into separate subsets
  based on the algebra of production and the structuring of relations between products. Instead of a single
  “flat” structure, it is proposed to decompose the product base into parts - to introduce a system of independent
  subsets of products. Parallel inference is implemented for individual subsets without loss of generality,
  while the search for possible alternatives is reduced. Each subset of productions has a special
  marker word, the value of which activates only one subset of productions. It is loaded into the operating
  part of a homogeneous computing system for parallel execution. Results. It is shown that quantitative
  estimates of the reduction in output time depend on the total number of productions, the number of subsets
  formed and their size. Simulation has shown that even the simplest decomposition into two subsets (one subset consists of 2 productions) gives a time gain of (1.07-1.52) times, proportional to the total number of
  productions. Conclusions. The created extended symbolic computing machine is the basis for the subsequent
  creation of the architecture of a homogeneous computing system with a combination of centralized
  and local control. This property allows computational units of a homogeneous operating part to work in
  parallel without excessive access to shared memory.

• ONTOLOGICAL APPROACH TO SOLVING THE WORKLOAD RELOCATION PROBLEM IN A DISTRIBUTED MONITORING SYSTEM WITH MOBILE COMPONENTS BASED ON A DISTRIBUTED LEDGER

  E.V. Melnik, I.B. Safronenkova, А.Y. Taranov

  2023-12-11

  Abstract ▼

  The paper considers the problems associated with the organization of the computing process
  in monitoring systems with mobile components based on a distributed ledger (DL), including the
  task of redistributing the computing load. The requirements for the functioning of modern distributed
  monitoring systems include the coordinated operation of nodes of the entire system belonging
  to various layers of the computing environment, including foggy and edge layers, which are highly
  dynamic. The joint use of DL technologies and mobile components as part of distributed monitoring
  systems makes it possible to expand the range of tasks solved by such systems, including due to
  the fact that it removes issues related to the synchronization of geographically distributed copies
  of data. However, with such an organization of a distributed system, it is necessary to take into account the following features of the computing environment: latency associated with data synchronization
  at DL nodes, changes in the geographical location of mobile components, limited
  onboard energy resources and high dynamism of the fog and edge layers. Previous studies have
  shown that in highly dynamic computing environments, the use of the search space reduction
  method based on ontological analysis is effective. For the correct operation of this method, it is
  necessary to develop an ontological model reflecting the features of the considered computing and
  communication environment, including DL and mobile components. In this paper a new ontological
  model of the functioning of a distributed monitoring system has been developed, taking into
  account the presence of mobile components and DL nodes. Production rules for placing computational
  load in foggy and edge layers have been developed and a software model has been implemented
  based on them, which allowed a number of computational experiments to be carried out.
  The results of experimental studies have demonstrated the effectiveness of the proposed approach
  and the adequacy of the developed ontological model.

• DEVELOPMENT OF A HIGH-PERFORMANCE METHOD FOR DETERMINING THE GEOMETRIC PARAMETERS OF OBJECTS IN THE IMAGE

  S.V. Onishchenko, E.V. Melnik, А.V. Kozlovsky

  2023-02-17

  Abstract ▼

  Currently, the development of various process automation systems is becoming more widely
  used every day in various fields and industries, the task of developing software methods for the
  corresponding automated systems remains urgent. One of the industries where the use and application
  of process automation systems is in demand is the field of non-contact measurement of objects
  and their parameters. As an example, the task of determining the geometric parameters of
  round timber stacked was chosen. In this regard, in this paper, methods were proposed for determining
  the geometric parameters of objects based on mathematical morphology operations, organized
  using the Canny detector and the Hough algorithm, and a method using a neural network
  approach based on the architecture of the YOLOv5 convolutional neural network. As a result of
  the conducted experimental studies, for the organization of which specially 3d-printed models of
  logs were used, it was found that the method based on the use of neural networks is more accurate
  than the method based on mathematical morphology. When solving the problem of counting the
  number of objects in the image, using the method based on the neural network approach, all objects
  located in the image were determined, whereas the method using mathematical morphology operations was able to determine only 13 of the 16 logs located, and I identified one false object,
  as a result of which the result error was about 19% for an image obtained from the Internet. When
  conducting an experiment on manufactured cylinder models, the method based on mathematical
  morphology operations showed unsatisfactory results. Another advantage of the method based on
  the neural network approach is the possibility of calculating the area of the ends of logs in the
  image and determining the volume of each of the logs located in the stack, as well as the total total
  volume of the entire pack of measured round timber.