
The relevance of the research is due to the need to develop effective methods for operational forecasting of grain crop yields. Traditional ground-based methods are time-consuming, while remote sensing data (RSD), especially hyperspectral imaging, provide a unique opportunity for non-invasive monitoring of crop conditions. However, the potential of RSD is often limited due to the high dimensionality of the data, multicollinearity, and insufficient knowledge of culture-specific spectral features. The aim of the work was to analyze the informative value of predictors and the possibilities of using machine learning methods to predict wheat yields using hyperspectral data. The following tasks have been solved to achieve it: collection and preparation of two heterogeneous datasets (from controlled test sites and from random points on an inhomogeneous field); development and implementation of an algorithm for automated search for new vegetation indices (VI) by iterating through combinations of spectral channels; conducting statistical data analysis to assess the significance of predictors; comparative analysis of the effectiveness of 11 feature selection methods and a variety of classification and regression algorithms. As a result of the automated search, more than 120 million combinations were processed and a number of new types were identified. Statistical analysis confirmed stronger direct correlations of yield with agrochemical indicators and multispectral indexes, while information from hyperspectral channels, although less explicit, turned out to be concentrated in certain areas of the spectrum. Forecasting experiments on these test sites have shown the critical importance of the feature selection stage. The best accuracy in the classification problem was achieved by the Gradient Boosting model with ANOVA selection of 10 features (Accuracy 0.77, F1-score 0.70), and in the regression problem by the support vector machine with RFE selection based on Random Forest (MAE 2.8, MAPE 16.7 %). The most significant features included the newly proposed indexes (VI1–VI9).
Bearings are critical components in modern machinery, and accurate fault diagnosis is essential for equipment reliability and maintenance cost reduction. Traditional methods relying on expert knowledge face limitations in complex scenarios. This paper presents a comparative study of classifier methods and Graph Neural Network approaches for bearing fault diagnosis. We systematically extract discriminative time-domain and frequency-domain features from vibration signals and develop multiple diagnostic models: for classifier-based approaches, we implement and optimize LightGBM, CatBoost, and gradient boosting classifiers via grid search; for graph-based approaches, we innovatively construct graph-structured datasets using K-nearest neighbor, radius, and path graph formulations, and design ChebyNet along with a modified Graph Convolutional Network (GCN) model incorporating a domain adaptation objective. Experimental results demonstrate that the proposed graph neural network methods achieve significantly higher accuracy (over 0.95) compared to traditional classifier approaches, with the GCN model incorporating Maximum Mean Discrepancy based domain adaptation yielding the best performance at 0.951 accuracy. This research provides insights into effective deep learning approaches for industrial fault diagnosis systems.
This paper presents the results of mathematical modeling of a two-dimensional field emission system based on an arbitrary number of blade-shaped field emitters. The emitters are located on a flat substrate. The anode is a plane parallel to the substrate plane. Since the surface of each emitter can be considered a corresponding equipotential surface, the influence of the emitters is replaced by the influence of charged filaments to calculate the electrostatic potential distribution throughout the entire emission system. The Poisson’s equation solution with the boundary conditions of the first kind is found as a series of eigenfunctions, with the coefficients of the series explicitly defined. All geometric dimensions of the system and the potentials at the electrodes are the parameters of the problem.
Performance testing of modern software systems is becoming increasingly demanding due to growing architectural complexity. Traditional methods focus on a limited set of critical configurations, which speeds up the process but reduces coverage and leaves potential performance bottlenecks undetected. To expand the coverage of system states, we introduce controlled perturbations into load estimation, which augments the data and improves the model’s ability to generalize to untested configurations. Using ensemble regression on tabular features and post-training model explanation, we accurately predict benchmark throughput and identify key factors influencing frequency selection and observed performance. In the case study of the Android kernel, the model achieves high accuracy, while Shapley Additive Explanations based interpretations align with the expected behavior of the scheduler and Dynamic Voltage and Frequency Scaling, suggesting a practical way to reduce testing effort and support engineering decisions in operating-system optimization.
The application of the exact penalty method to the problem of finding a cost-optimal trajectory is considered. To minimize the exact penalty function, the apparatus of constructive nonsmooth analysis is employed. It is shown that the application of the exact penalty method in the original functional formulation leads to the generation of successive approximations in the form of non-elementary functions. Although this fact is theoretically expected, this work is the first to investigate its practical consequences: it is demonstrated that as early as the second iteration, the computational complexity increases to such an extent that the analytical implementation of the algorithm becomes infeasible (a slowdown of more than 10 000 times compared to the Ritz method). To overcome these difficulties, a grid-based adaptation of the method is proposed, which allows obtaining a solution in a reasonable time. The obtained results may be useful for researchers planning to apply the exact penalty method to problems in the fields of calculus of variations and control theory, as well as for solving trajectory optimization and route planning problems.