Intelligent Systems

At the stage of development of transport logistics in the conditions of the data economy, the processes of information transmission to railway, automobile and aviation vehicles that carry out the movement of goods or products at distances in a transport corridor equipped with radio access points are being investigated. The task of ensuring continuity of communication with vehicles over the distances of the transport corridor has been set and solved, in which there are differences in the parameters of access to the data transmission network in the geographical zones of this corridor. New information and logical models for intelligent management telecommunications have been developed – data transmission from the control point to the vehicle. The novelty of the models lies in ensuring continuous transmission of information when passing a vehicle in neighboring distances, which differ in the parameters of access to the data transmission network. To implement novel models, it is recommended to apply a method consisting in the implementation of these models in well–known technical solutions - in digital platforms for supporting organizational systems processes. The positive effect of the use of novel models is to ensure continuous communication with vehicles and, as a result, to increase stability when remotely controlling the movement of these vehicles along the distances of the transport corridor.

This article is devoted to the comparison of algorithms for estimating the position and heading angle of the autonomous underwater vehicle (AUV) relative to the stationary landing platform (SLP) using a short-range hydroacoustic system. The problem is mathematically formulated using the Bayesian framework. Three algorithms are compared. One of them involves non-inertial processing, while the other two take into account the accumulation of measurements during the AUVs movement. One of the two latter algorithms is implemented in a recursive scheme: the recursive iterative batch linearized smoother (RI-BLS), while the other one follows a non-recursive scheme: an algorithm based on factor-graph optimization (FGO). Simulation results comparing the root mean square errors (RMSE) for the AUV’s position and heading angle estimates are presented, evaluating the performance of the algorithms. These RMSEs are also compared with the Cramer-Rao lower bound.

The paper presents QuadTree-Based Graph Convolutional Networks (QGCNs), which use a quadtree probability model that is first applied to small object segmentation in synthetic aperture radar (SAR) images for the case of insufficient training data. QGCNs are ensemble architectures that consist of a pre-trained encoder that forms SAR image pixel features and a neural quadtree, which is a graph convolutional network, a neural network analogue of a quadtree probability model. The graph network also contains a special branch pruning block and upscaled image features for highlighting similarities between pixels at different spatial resolutions. QGCNs with U-Net as an encoder are used to segment small objects in a few real SAR images (from the Sentinel-1 and HRSID datasets). QGCNs show higher quality in the segmentation of small objects in both multi-class and two-class problems. The difference between QGCNs and U-Net in Recall value for small objects in a multi-class problem is up to \(10.32\%\). The increase in \(F_1\)-score for QGCNs in a two-class problem is up to \(3.59\%\).

This study focuses on the analysis of the process of adjusting weights in neural networks, highlighting key aspects such as the influence of the error vector and the activity of previous layers on weight changes. The work proposes methods that allow for effective management of activity increment vectors and ensures their alignment along the error vector direction. Special attention is given to threshold-adjusted mapping and the algorithm for vector orthogonalization, which contribute to improving approximation accuracy and increasing learning speed, which is important for control tasks. The results of the study emphasize the importance of accurately determining error vectors and suggest promising directions for further research in the field of machine learning.

This paper considers the influence of different color space resampling methods on the training quality of neural networks of fully connected and convolutional architectures. It also studies the use of the Haar and Daubechies D4 wavelet transforms and wavelet filtering before space resampling in order to improve the quality of neural networks after color space resampling. The resampling methods considered are a bitwise shift by 3 bits used to switch from the RGB format to the R5G5B5 format and a strong compression method with splitting the color space into two, three and four intervals, allowing storing the pixel color value in two bits. A dataset of black-and-white images of handwritten digits is used to test the influence of the resampling algorithms on the quality of neural networks. At the end of this paper, the accuracy metric values are given for two architectures using two types of wavelet filtering and eighteen types of resampling.

This paper is devoted to the application of machine learning methods to the problem of identifying parameters of dynamic systems. The system under consideration is the Twin Rotor MIMO System mechatronic bench, which is a simplified model of a helicopter with two degrees of freedom. The dynamics of motion are described by a system of nonlinear differential equations, which does not allow the use of classical methods based on parameterization of the model in the form of linear regression. To solve the problem, the following methods were used: multilayer perceptron, polynomial regression, Bayesian regression and ensemble of models. Conducted exploratory data analysis to evaluate the quality of the dataset. A comparative analysis and evaluation of the obtained results are presented.

This study investigates a production (s, S) inventory structure where clients arrive based on Poisson process and the commodities are perishable. The length of time between each item’s manufacture is exponentially distributed. The service time in this single server system is also exponentially distributed. If the inventory is not empty and the server is available, a customer is served upon arrival. When a newly arriving customer discovers that the only server of the system is busy, they have two options: either leave the system or go to a waiting area known as the orbit with unlimited capacity. Each customer in the orbit uses an exponentially distributed time interval to attempt a second visit to the service facility. When the stock level is above or equal to ‘s’, the service rate is increased, to prevent loss due to deterioration. Using the Matrix Analytic Method, a solution to the problem is obtained using an algorithmic approach. Various performance indicators of the system modelled are derived. The effects of various characteristics on these measures are investigated quantitatively. The optimal profit is determined by constructing a suitable profit function.

The work considers the problem of planning and optimizing the trajectory of a mobile 2D robot processing a flat area limited by arbitrary closed boundaries. The desired trajectory must be optimized according to three criteria: minimum processing time, fuel consumption (energy) and processing resource consumption (liquid, seeds, etc.). In addition, the desired trajectory must have a parameterization that allows for further optimization and application in artificial intelligence systems. The main idea used is ternary coding of the neighborhood of a point and algorithms that exploit the connectivity of spatial regions on a lattice. The trajectory builder creates a partially optimized polyline, the construction of which generates its parameters, which are then used for optimization by a genetic algorithm. To use a genetic algorithm, parameters of different power are reduced to a system of parameters of the same power, equal to 3. The experimental construction and optimization of the trajectory confirmed the viability of the proposed algorithm and the idea of trajectory planning.

The article addresses the problem of detecting and isolating sensor and actuator faults in a class of MIMO (multiple-input, multiple-output) systems characterized by an arbitrary relative degree. The proposed solution is based on the design of observers that remain invariant to the input signal. Simulation results are provided to demonstrate the effectiveness of the method. Additionally, the approach ensures robustness against disturbances.

The article is devoted to the construction of concept models of information application. Such models shall be used as the basis for the construction of mathematical models of information application. Human or actor (agent) activity considered for application of information. Role of information in human activity and role of information application for the activity considered. As a result, it is shown that such models currently have insufficient quantity and quality. As an example of such a situation, Robert Solow’s paradox persistence and its latest manifestations are considered, and existing information quality principles are discussed. It is shown that to overcome the gap between the required models and the available ones, the complex of mathematical models of information application shall be proposed. Suggested results made it possible to model information application for activity with mathematical models and methods. Some kinds of application of information cases are discussed on the basis of the proposed concepts and schemas. The information application is introduced as a sequence of information applications for activity problems and to make decisions in practice.

This paper presents a control architecture for aircraft, designed to enhance positioning accuracy and reliability during taxiing and runway operations. The system includes a data monitoring module that activates a position prediction mechanism using a Convolutional Neural Network (CNN) when GPS fails. Additionally, an intelligent control module based on transformer architecture manages aircraft movement. The architecture incorporates a Decision Transformer for reinforcement learning, refining control policies via supervised learning and Proximal Policy Optimization (PPO). An obstacle detection module, utilizing YOLOv7, ensures rapid and accurate object detection during taxiing. Experimental results validate the system’s efficiently across various weather and operational conditions, maintaining minimal deviation from the runway centerline and handling abnormal scenarios like engine or brake failures. The developed software and algorithmic support, tested in the Xplane-11 simulation environment, provide a robust solution aircraft control, ensuring safe and precise operations under diverse conditions.

In contemporary industrial production, the trend toward automation is pervasive across all sectors, highlighting the necessity for effective tools to facilitate the development of automated systems. This article explores the implementation of the AutoML approach to streamline the creation of machine learning models, particularly for applications in automatic monitoring. The AutoML approach offers significant advantages in the field of machine learning by automating the process of model selection, tuning, and evaluation. By reducing the reliance on extensive domain expertise, AutoML democratizes access to advanced analytical tools, enabling non-experts to effectively develop models. This approach enhances efficiency, as it streamlines workflows and minimizes the time required to deploy models in practical applications. The paper proposes the developed software plat-form AutoGenNet. It is implemented within the AutoML paradigm and utilises the No-Code development concept. This concept effectively abstracts the complexities associated with model creation and training, thereby lowering the entry barrier for users without extensive technical knowledge. By utilising No-Code methodologies, the platform not only simplifies user interaction but also increases accessibility. Furthermore, the AutoGenNet platform includes a mechanism for automatically generating software wrappers, facilitating efficient operation of trained models. This comprehensive integration allows for the effective application of the AutoML approach in automating the generation and training processes of the machine learning model. As a result, the system significantly accelerates and simplifies the resolution of automatic monitoring tasks utilizing machine learning methodologies. In addition, the developed system is designed with scalability in mind, allowing it to be adapted in the future for automated generation of various models of other machine learning architectures. This flexibility opens up new possibilities for solving a variety of practical problems in monitoring applications, thus extending the usefulness of machine learning in industrial contexts.

Digital signatures are essential for maintaining the authenticity and integrity of communications within resource-constrained devices like microcontrollers, which operate with limited processing and memory. This study seeks to determine the most efficient and secure digital signature algorithm suitable for microcontrollers used in embedded systems and IoT devices. Consequently, the research examines the implementation of three digital signature algorithms—RSA, ECDSA, and EdDSA—within these specific environments. The methodology involved mathematical analysis and the implementation of RSA, ECDSA, and EdDSA algorithms using the Python programming language, followed by performance testing in terms of time and memory usage during key generation, signing, and verification processes. The results indicate that EdDSA offers the best balance of efficiency and resource usage, making it the most suitable for microcontroller-based systems. ECDSA also demonstrates strong performance but is slightly less efficient than EdDSA, and it is known to suffer from certain security vulnerabilities, such as issues related to nonce generation. RSA, while providing quick verification, is the most resource-intensive and thus less ideal for constrained environments.

This study provides practical insights into the trade-offs between different digital signature algorithms, offering clear guidance on selecting the most appropriate algorithm for secure and efficient operation in resource-constrained systems. These findings are crucial for developers aiming to implement secure and efficient digital signature algorithms in microcontroller systems, where resources are inherently limited.

The modern code review is actively used and developed in connection with standard at the present time practice for application of the distributed software development in IT-companies. In this kind of review, a special software tool, supporting the entire process from submitting a change request by a code developer to approving this request by code reviewers with updating the project repository or rejecting the request, is utilized. A crucial point of the review is the procedure of assigning code reviewers to change requests submitted. On how well the assignment will be done depends not only the code review result but, subsequently, also the IT-project itself: its timing and software development quality. The code reviewers’ assignment task becomes even more complex when projects are large-scale and can involve hundreds and thousands of developers. In them, numerous requests from developers are created concomitantly. Regardless of the fact that at the current time, a question of assigning code reviewers is under active study, the subject of optimal selection (when the problem solution is a list of code reviewers on each submitted change request to achieve maximum total compliance) and accounting different common reviewers’ features, for example, the experience as a developer and experience as a reviewer, is poorly explored. The suggested approach to decision-making support in assigning code reviewers is developed to resolve the issue. The operability of this approach is confirmed by the experiment using the generated dataset.

This research investigates a k-out-of-n reliability system in which components are prone to failure, and a single unreliable server is responsible for their repair. During idle periods, the server utilises its time to serve external customer requests. Repair operations for internal components commence only after N such failures have occurred. To maintain the system’s reliability during the server’s engagement with external customers, an N-policy is enforced. If N internal failures take place while external service is ongoing, the server preempts the external service to address internal issues. The failure times of components and the arrival of failed external customers are assumed to follow Poisson distributions. The service times for both external requests and system repairs follow a phase-type distribution, while the server’s breakdown and repair duration are also modelled using exponential distributions. The Matrix Analytic Method is employed to analyse the system’s steady-state behaviour and stability. Additionally, a cost function is defined to facilitate the numerical optimisation of the N-policy threshold.

The well-known mathematical model SEIR of the epidemic process is considered. The work is devoted to solving the optimisation problem in the scenario of countering the spread of the infectious disease. The purpose of the study: to find the cost-optimal control effect in a scenario of three applied measures: mobility restrictions, vaccination and preventive prophylactic measures, provided that the proportion of infected persons does not exceed the established value.

The SEIR model was used to simulate the spread of infection. The decision was to solve the model approximately, not numerically, to make optimisation easier. The relevant propositions were taken: zero mortality and the models for considered factors (mobility restrictions, vaccination and preventive prophylactic influences) on the predicted state. The task of optimising the introduced measures is the task of finding such a set of model parameters that depend on the control actions, in which the minimum cost of measures is achieved while observing the condition of limiting the proportion of infected persons. The solution was carried out by the internal point method, the implementation of which turned out to be computationally efficient after a piecewise approximate solution for the zero-mortality SEIR model was found.

Using the found approximate solution, an algorithm for searching for the optimal combination of infection containment measures with a limit on the number of infected is constructed. The algorithm found is a general way to optimise infection containment measures, provided that it develops according to the SEIR model with no mortality.

In this paper the problem of sensorless control for permanent magnet synchronous motors (PMSMs) with uncertain parameters is considered. The designed approach is based on adaptive flux and position observer with fixed-time convergence and uses classical field-oriented controller. It is assumed, that stator currents and control voltages are measurable signals and the only known motor parameter is the stator inductance. Presented estimation algorithm is based on Kreisselmeier’s dynamic extension of the regressor with suitable mixing and ensures fixed-time estimation of the motor fluxes, position and stator resistance. The efficiency of the proposed solution is compared with state observer that uses dynamic regressor extension and mixing algorithm. The improvement of transient behavior is demonstrated via simulation results.

Nowadays, a mobile object groups management is a relevant scientific problem. The structural complexity, heterogeneity, and geographical distribution, along with the complex criteria of their efficiency estimation make the efficiency improvement of such systems a computationally hard optimization problem. This makes it inexpedient to solve such computational problems by means of analytical or mathematical programming methods usage. Besides, mobile object systems are dynamic in terms of constraints and criteria of their models. So, the question of the resources, which are needed for such computationally hard problem solving, emerges. The aim of this research is to reduce the time required to solve discrete optimization problems in a distributed heterogeneous computing environment while maintaining the solution accuracy level. The novelty of the method proposed in this paper is the technique of distributed optimization problem solving for dynamic computing environments, which is implemented by means of efficient distribution of metaheuristics instances through the available nodes.

The paper considers the problem of developing artificial intelligence tools in the field of project management. Models for predicting missed deadlines within the framework of national project checkpoints are prepared using machine learning methods. For this purpose, anonymized and normalized data of the monitoring system for 2022-2024 with a total volume of 4,783 records are used. The task is to perform binary classification. The data for 2022-2023 were used to train models (20% of data withheld as test data) while 2024 year data were used to validate the resulting models. Cross-validation and oversampling methods are used to eliminate class imbalance. The results indicate a difference in data between the years. Because of this, the use of models trained on the first two years data for predicting based on the third year data shows slightly lower scores. Numerical quality scores of the models are presented and discussed. Possible directions for future research are noted.

The article proposes an approach to improve the interaction between humans and a collaborative robotic system (CRS) based on multimodal emotion recognition. Modules for speech, text, video and voice analysis are presented, as well as methods for aggregating the results of various modalities. The developed architecture of the multimodal system allows you to adapt to changes in the emotional state of the operator in real time. During the experiments, the proposed algorithm was evaluated using real industrial data. Special attention is paid to the training of models, accuracy metrics and their integration into the overall architecture of the collaborative robotic systems. The results of the study demonstrate the prospects of using multimodal approaches in industrial robotics. Experiments have shown that the proposed approach demonstrates high accuracy in recognizing emotions, surpassing traditional methods. The results obtained confirm the possibility of using the developed system in collaborative robotic systems to improve adaptability and interaction with operators in dynamic production conditions. Further development of the method will be aimed at increasing the system’s resistance to noise and various data variations, as well as expanding its application in other areas.

Machine learning (ML) models are increasingly used for time-series analysis to predict the behavior of complex dynamic systems, such as automatic process control systems and robotics. However, these models are vulnerable to various security threats, especially in adversarial environments. One prominent attack method is the Fast Gradient Sign Method (FGSM), which perturbs input data in a way that maximizes the model’s prediction error. This paper explores the unique vulnerabilities of ML models applied to time-series data, focusing on adversarial attacks (demonstrating influence of FGSM attack). Additionally, we discuss defense mechanisms to mitigate these risks and ensure the robustness of these models in safety-critical applications.

An online learning evolution and systems of its management exploitation have led to the possibility of accumulating data on the education process. The accumulated data provide a basis for analytics and subsequent decision-making of online teachers and university administration. Researchers utilize educational data mining for different purposes by means of searching templates in these data. A field connected with controlling quality of fulfillment of free response assignments has remained largely untouched, though. Decision support for online teachers, implementing this procedure, allows to reduce the influence of the human factor, associated with over-loading and time constraints, and enhance the control quality. To carry out such support for online teachers, an approach and intelligent decision support system, developed with the usage of mining data being stored in a learning management system when controlling quality of fulfillment of free response assignments, are proposed in the paper. Based on experiment results, it was discovered that the system detects errors in performed free response assignments missed by online teachers, which proves its efficiency and usefulness for education process.

The principle of functional voxel construction of complex computational processes is given on the example of modeling the R-function of the union or the intersection of the domains of two functions. The basics of arithmetic operations on the local geometric characteristics describing the components of a homogeneous unit vector of a local function are analyzed. The principle of de-normalization of such components for use in arithmetic operations that make up the R-function is demonstrated. The modeling of the scene in the form of a layout of concentric objects and a local function of target description by the surface of the funnel at a specified point is considered. The algorithm of dynamic formation of the final local function of the union of the surface of the funnel with the surface of the scene at the current point is considered. Based on the final local function, the components of the vector of the gradient movement towards a given target are determined.

When considering a path given as a sequence of GPS points, this path is often noisy and erratic because of poor accuracy of GPS receiver. Since the points are often recorded periodically, it seems appropriate to restore the smoothness of the trajectory by window-filtering the sequence in order to align them in order. However, to smooth an actual trajectory, it is typically required to take wide window for filtering. Wide windows introduce a substantial value of trajectory shift on turns towards the center of the turn. Current article proposes a simple yet efficient method for restoring the in-turn displacement for effective use of window filtering without involving knowledge about actual motion of the GPS receiver.