Propulsion system testing plays a critical role in the development of experimental sounding rockets by providing experimental validation, reducing engineering uncertainty, and supporting informed design decisions. While test campaigns require additional infrastructure and operational resources, they enable improvements that are often unattainable through analytical methods alone. This paper presents both theoretical and practical aspects of propulsion system testing for sounding rockets, including measurement techniques for thrust, pressure, temperature, and propellant mass, together with selected solutions implemented in the propulsion system test stand developed by the student research group PWr in Space.
Machine learning (ML) is increasingly recognized as a powerful tool for scientific discovery, yet practical guidance for working with small to medium-sized datasets is limited. This article fills that gap by presenting lessons drawn directly from the author’s hands-on experience across natural and social sciences, including biomedical signal analysis, survey research, and behavioral prediction. Unlike a literature review, it reflects real-world applications, illustrating what works in practice and what pitfalls to avoid. The study outlines a structured ML workflow emphasizing careful data preparation, model selection, rigorous validation, and interpretability. Both shallow and deep models are considered, with advanced techniques such as SHAP used to reveal how models make decisions and extract meaningful insights. Results highlight that effective ML depends less on algorithmic complexity and more on disciplined methodology. Interpretable models, integration with domain knowledge, and thoughtful validation often outperform more sophisticated alternatives on modest datasets. Common challenges—including data leakage, default-model overreliance, and misconceptions of ML as an automatic solution—are addressed with practical examples. This article serves as a guide for researchers who want to apply ML responsibly and effectively, demonstrating how real-world experience can transform small or imperfect datasets into scientifically meaningful insights.
The paper analyses the use of artificial intelligence (AI) in the risk and reliability management of IT systems in the context of the DevSecOps approach, which integrates security into the software lifecycle. In response to the growing number of cyber incidents and the limitations of traditional methods, specific AI tools and algorithms used in the automation of security tests, IT infrastructure monitoring and failure prediction were presented. The research was based on a review of 42 scientific publications and the analysis of empirical data from five IT organizations, using quantitative and qualitative methods (surveys, interviews, case studies). The results indicate that the integration of AI in DevSecOps increases the effectiveness of threat detection by 35%, reduces incident response time by 40%, and improves the reliability of IT systems. Key challenges such as data quality, model interpretability, and regulatory compliance were also identified. The article formulates recommendations for DevSecOps teams and technology decision-makers and indicates directions for further research in the field of ethics, interoperability, and auditability of AI systems.
This article presents the current capabilities for measuring the accuracy of selected IBM Q-class quantum architectures for specific stages and tasks of the recommendation process in hybrid classical-quantum recommendation systems. The main motivations for this research undoubtedly include the increased use of quantum technologies to accelerate the computational process in recommendation systems, as well as the need to assess the utility of available quantum processors for tasks requiring high accuracy. A comparison of the implementation of complex multi-qubit systems in a quantum simulator with that on a real quantum computer is discussed to demonstrate the measurement accuracy of currently available quantum architectures and the error rate compared to simulations of near-ideal systems. The literature lacks consistent comparative analyses that compare real implementations of multi-qubit systems on real quantum computers with their ideal simulations. The existing works mainly focus on model examples or theoretical analyses, which leaves a gap regarding the estimation of the actual error rate and stability of the performed calculations in recommendation applications. This motivated the presentation of important aspects of error generation and the identification of their causes.
This study investigates the use of machine learning methods to assess and predict the level of athletes’ preparation for e-cycling competitions based on their training data. A three-month dataset of activities collected from selected competitors of the Polish Open E-cycling Championships was processed to extract key performance indicators describing individual training cycles. Three supervised learning models – linear regression, decision trees, and artificial neural networks – were evaluated, achieving average classification accuracy of approximately 0.6. Additionally, an unsupervised k-means clustering approach was applied to identify natural groupings of athletes based on multidimensional training characteristics. The findings indicate that the constructed dataset enables reasonable prediction of preparation level; however, broader and more diverse data, including contextual factors such as well-being, nutrition, and environmental conditions, may be required to significantly improve model performance. The results highlight both the potential and the limitations of applying machine learning techniques to support planning and evaluation of training in e-cycling.
The rapid growth of Internet of Things (IoT) devices in smart cities produces substantial and heterogeneous network traffic, creating significant challenges for security and management. Accurate device classification is crucial for protecting networks against threats, such as malicious devices, and for facilitating efficient resource allocation and management. Although various classification methods have been proposed, a comprehensive review evaluating their effectiveness in addressing the specific challenges of smart cities, such as large-scale device heterogeneity and extensive encrypted traffic, remains lacking. This paper provides a systematic review of the literature, analyzing and comparing IoT device classification methods that utilize network traffic data. The methodology emphasizes two primary aspects: feature extraction techniques (including packet-level, flow-level, and automated approaches) and the classification algorithms employed (supervised, unsupervised, and deep learning methods). The analysis shows that packet-based methods achieve high precision but face limitations with encrypted traffic and scalability. In contrast, flow-based and deep learning approaches exhibit greater adaptability. Machine learning (ML) and deep learning (DL) algorithms consistently demonstrate strong performance, with reported accuracy rates reaching up to 99%. The review identifies several critical future research directions for smart city applications, including experimental validation, integration with edge computing, and the development of hybrid classification models.
The high competitiveness and rapid pace of development of the manufacturing industry requires companies to constantly improve, both in terms of efficiency and quality of production. It is these parameters that directly affect not only the reputation of the company, but also the costs it incurs, ultimately determining the price of the product on the market. Both production efficiency and product quality are affected by the speed and effectiveness of detecting manufacturing defects. This paper presents a study to evaluate the effectiveness of detecting defective production quality in real time, using Convolutional Neural Networks (CNN), You Only Loon Once v8 (YOLOv8), and Faster R-CNN with ResNet-101 backbone. The study was carried out on the example of data for assembly lines in the automotive industry, supported by advanced techniques of pre-processing, data augmentation and knowledge transfer. The high competitiveness and rapid pace of development of the manufacturing industry requires companies to constantly improve, both in terms of efficiency and quality of production. It is these parameters that directly affect not only the reputation of the company, but also the costs it incurs, ultimately determining the price of the product on the market. Both production efficiency and product quality are affected by the speed and effectiveness of detecting manufacturing defects. This paper presents a study to evaluate the effectiveness of detecting defective production quality in real time, using Convolutional Neural Networks (CNN), You Only Loon Once v8 (YOLOv8), and Faster R-CNN with ResNet-101 backbone. The study was carried out on the example of data for assembly lines in the automotive industry, supported by advanced techniques of pre-processing, data augmentation and knowledge transfer. The results of the research made it possible to assess the potential of the analyzed solutions in detecting defects. The highest average precision (mAP₅₀) was achieved using YOLOv8. By using Faster R-CNN with the ResNet-101 backbone, higher precision was achieved only for small defects, but at a lower speed. When using the standard CNN neural network, the average precision was the lowest, while at the same time not providing spatial localization capability. The main barriers to implementation include computational requirements, as well as integration with programmable logic controllers (PLCs).
One of the primary problems in our society is skin diseases, due to their severe effects on both the body and the mental health of individuals. Early detection and diagnosis of these diseases can have a significant impact on the success of treatments. The skills and experience of specialists play a crucial role in determining effective methods for diagnosing and treating skin lesions. In the early stages, advanced techniques were developed to automatically test skin for diseases. In recent years, skin diseases have increasingly been diagnosed with the help of artificial intelligence through machine learning algorithms, which are trained on vast amounts of data already available in the healthcare sector. In this research report, we comprehensively investigated previous studies on the use of machine learning in skin disease classification. Skin diseases were successfully classified in several studies with varying levels of diagnostic accuracy. Some studies used image processing and feature extraction for this purpose, while others focused on specific types of skin diseases by utilizing clinical features obtained from tissue analyses of the affected areas. This research report concludes that using image processing methods, accuracy ranged between 50% and 100%. Another method, involving tissue analysis, achieved an excellent accuracy rate of 94% or higher. The findings present a review of the related studies conducted in the literature and focus on the recent research gaps.
This paper presents a conceptual multilayer model designed for the analysis and simulation of spatial conflicts, incorporating the interdependencies between infrastructural layers and socio -spatial dynamics. The model employs a cellular automata (CA) approach, enabling the exploration of local interactions within complex spatial systems. Each layer represents a distinct aspect of the system: society, technical infrastructure, and the spatial- environmental context. The paper discusses the mechanisms of information exchange between layers, conflict escalation processes, and example scenarios derived from dependency network analysis. Conceptual results indicate that the multilayer approach provides a more realistic representation of crisis dynamics and supports the development of more effective strategies for spatial planning and response management.
The transition to Industry 4.0 has continued to evolutionise maintenance strategies, shifting from reactive approaches toward new, data-driven predictive maintenance systems. This study evaluates three machine learning approaches—Random Forest, XGBoost, and Long Short-Term Memory networks—for equipment failure prediction using live sensor data from rotating machinery. The study is based on feature engineering from multi-sensor time-series data and time-dependent validation protocols. The research demonstrates that mixed methods achieve superior performance in binary failure classification (Random Forest with 94.2% accuracy, XGBoost – 93.7%), while Long Short-Term Memory networks perform particularly well in remaining useful life prediction with RMSE values 23% lower than traditional approaches. The analysis reveals that model selection should be guided by specific maintenance objectives: immediate failure detection favors mixed methods, whereas lifecycle planning benefits from deep learning architectures. The study identifies key implementation barriers including data quality, computational requirements, and integration with Manufacturing Execution Systems. Comparative analysis demonstrates significant economic viability with 15-month payback periods and >80% annual ROI in high-downtime environments.
Progressive urbanization and increased air pollution emissions pose a significant challenge for modern safety engineering, especially in terms of forecasting and mitigating environmental risks1. This paper presents a machine learning-based approach to modeling PM2.5 particulate matter concentrations using open environmental and meteorological data from the OpenWeatherMap platform. The aim of the study was to develop a model to support decision-making in environmental safety systems through early detection of potential air pollution episodes. The study used a Random Forest model with parameter optimization and cross-validation, as well as time feature transformations to account for the cyclical nature of atmospheric phenomena. The analysis of the results showed a high correlation between the predictions and the actual PM2.5 concentrations, with a coefficient of determination R² above 0.8 in most of the analyzed time intervals. The results confirm the effectiveness of the proposed approach in identifying trends and environmental anomalies that may pose a threat to public health. The developed model can be a component of intelligent environmental safety management systems2 and a basis for further research on the integration of artificial intelligence with IoT infrastructure and urban air quality monitoring systems.
This paper presents the concept of an agent-based, distributed web application utilizing artificial intelligence (AI) for forecasting startup profitability. The proposed system integrates key success factors (KSFs) to evaluate financial and operational performance across different stages of a startup’s life cycle. The application employs autonomous agents that collect, process, and analyze multidimensional data to generate profitability forecasts and recommendations. The multi-agent architecture ensures scalability, adaptability, and resilience through asynchronous communication and self-organizing behaviours. Seven types of agents are defined performing specialized tasks to enable collaborative learning and dynamic decision-making. The system design allows further extensibility by adding new agents. This concept provides a foundation for the development of intelligent decision-support tools that assist entrepreneurs and investors in making data-driven strategic choices throughout a startup’s lifecycle.
Modern industrial environments require reliable, low-latency communication, often exceeding the capabilities of public mobile networks. Private 5G networks address these needs by enabling full control over transmission parameters and security. This paper presents the planning, testing and optimisation of such a network and evaluates its performance through call setup and data transmission measurements. The results confirm that precise base station configuration and continuous optimisation are essential to maintaining high service quality in industrial applications.
This study explores the application of machine learning methods to predict football (soccer) match outcomes. Football, as a highly dynamic and data-rich sport, provides a valuable source of information for predictive modeling. The research focuses on evaluating and comparing the performance of several machine learning algorithms: a naïve baseline model, logistic regression, random forest, XGBoost, and an artificial neural network. The dataset used for training and testing consists of historical match statistics, team performance indicators, and situational variables such as home advantage. Feature engineering and data preprocessing steps, including normalization and handling of missing data, were applied to improve model performance and generalizability. Each model was assessed using standard evaluation metrics such as accuracy, precision, recall, and F1-score. The results indicate that while simple models like logistic regression provide solid baseline performance, ensemble methods such as random forest and XGBoost achieve higher predictive accuracy. The artificial neural network, although more computationally demanding, shows promising results in capturing complex, nonlinear relationships between match variables. The study highlights the challenges of modeling football outcomes, such as the inherent randomness of the sport and the influence of non-quantifiable factors, but demonstrates the potential of machine learning in enhancing predictive analytics in sports.
Forecasting production demand is one of the key elements of effective supply chain management in small and medium-sized enterprises. The paper presents a comparative analysis of different neural network architectures used in production demand forecasting. The study includes the evaluation of neural networks: unidirectional (MLP), recursive (RNN), long-term memory (LSTM), recursive units (GRU), convolutional networks (CNNs) and hybrid models combining multiple architectures. The analysis showed that the choice of the appropriate network architecture depends on the characteristics of the time data, the degree of complexity of demand patterns and the available computing resources. LSTM networks show high performance in modeling long-term time dependencies, while CNN-BiLSTM hybrid models offer the best results in the context of multivariate time series. The article provides recommendations for choosing optimal architectures for specific production scenarios.
In software architecture design, proper attention is often not paid to the relationships between elements across diagrams. However, IT designers, more or less consciously, introduce elements representing the same object instance in different diagrams with nearly identical names. These relationships are called consistency rules and are typically not supported in current software modeling tools. A software architecture without consistency rules is simply an unrelated set of diagrams. Failure to apply consistency rules leads to countless inconsistencies in software architecture. This article presents a mathematical proof demonstrating that applying consistency rules primarily increases the information content of a software architecture. Therefore, the concepts for improving the readability and orderliness of software architecture using consistency rules also have a strong mathematical basis. This article demonstrates the reduction of information entropy (information uncertainty) when applying consistency rules in the design of software architecture of IT systems. It has therefore been proven that labeling selected elements in different diagrams with similar names, indicating the use of consistency rules, is a very effective way to increase the information content of a software architecture while simultaneously improving its orderliness and readability. Considering the constantly increasing quality requirements placed on IT systems, it seems that without IT specialists paying more attention to consistency rules in software development, it will not be possible to significantly improve the current stagnation in software development methods.
The fast rise of digital data brings both chances and problems for law enforcement and forensic investigators. The large amount of digital evidence can give useful information about crimes, but the high volume and complexity of this data make manual analysis hard. Machine learning (ML) algorithms have become a strong tool for automating and improving digital forensic investigations, providing benefits like better accuracy, efficiency, and scalability. This paper looks into how well ML algorithms work in finding and analyzing digital evidence, examining their ability to simplify forensic processes and discover hidden patterns in complicated datasets.
Diverse sources of radio signal interference and the complex topology of multi-level indoor structures often impede reliable position tracking. Therefore, a solution based on a Pedestrian Dead-Reckoning (PDR) navigation module utilizing high-frequency data from Inertial Measurement Unit (IMU), as well as two complementary technologies: Bluetooth Low Energy (BLE) – employing RSSI-based distance estimation and Ultra-Wideband (UWB), which relies on Time-of-Flight (ToF) signal, was proposed. The continuous, high-frequency motion tracking provided by the IMU was periodically corrected using absolute, though inherently noisy, radio-based measurements. This fusion strategy allowed for the mitigation of drift accumulation, yielding a trajectory estimate that remains both locally smooth over short intervals and globally accurate over extended durations. An Extended Kalman Filter (EKF) served as the fusion mechanism, integrating relative motion estimates from the PDR with absolute positional updates to perform a series of controlled experiments under diverse conditions. The analysis of the collected data revealed that the standalone PDR system exhibited a substantial drift error on average between 10% and 15% of the total traversed distance. Sensor fusion with BLE measurements significantly reduced this error, achieving a localization accuracy of Root Mean Square Error (RMSE) = 0.98 m. Under analogous conditions, the UWB-based system demonstrated a decisive advantage, reaching an accuracy level of RMSE = 0.21 m, corresponding to nearly a fivefold improvement compared to BLE. The experimental results confirm that UWB, when combined with sensor fusion frameworks, provides a balance among positional accuracy and robustness to environmental disturbances, making it effective for precise indoor localization.
Optimization in job-shop production systems is a very complex task, requiring simultaneous analysis of many factors. This article presents just this type of production in which the problem was to schedule job-shop tasks with significant setup times. In this case, the use of reinforcement learning (RL) was proposed, with particular emphasis on algorithm evaluation: Q-learning, actor-critic algorithms, deep neural networks (DQN), as well as advanced neural graph-based architectures (GNN). As a result of the analysis, it was shown that methods that are based on deep reinforcement learning (DRL) allow to achieve better results than traditional distracting heuristics and metaheuristic algorithms. Neural networks, which are based on heterogeneous graphs, are particularly useful for modeling relationships between machines and operations in the face of dynamically occurring changes in the production schedule. In this article, the methods described are analyzed in the context of production in small and medium-sized enterprises (SMEs), for which intensive technological development is a challenge. Ways of implementing RL solutions for such enterprises were also proposed.
This article presents a comparative analysis of an implemented modern Intrusion Detection System (IDS) based on artificial intelligence algorithms and traditional solutions, with particular emphasis on the Snort platform. The study includes a comprehensive evaluation of the effectiveness of various IDS/IPS systems in test environments, analyzing their ability to detect and block network and application-level attacks. The model based on the RandomForestClassifier algorithm achieved 99% accuracy, demonstrating high effectiveness in detecting UDP flood attacks. The comparison with Snort 3 system reveals significant differences in detection methods, operational efficiency, and practical aspects of implementation. The research results indicate the complementary nature of AI-based systems and traditional signature-based methods, suggesting the optimal use of both approaches depending on the environment’s characteristics and security requirements.
