Methods for classifying audio signals received from several simultaneously active sources and with partially overlapping features are being investigated. Real audio recordings often contain sounds from multiple sources, which significantly complicates the task of automatic recognition and reduces the accuracy of standard models trained on single-component signals. The purpose of the study is to evaluate the effectiveness of various classification scenarios for multicomponent audio signals. The ResNet18, ResNet34, and ResNet50 architectures are used in the experiments. Models trained on single-component audio signals and tested on multicomponent ones using classical spectral filters and the Demucs neural network separator, as well as models trained directly on multicomponent signals, are considered.

Multicomponent signal training provides the highest classification accuracy, up to 88.5% on test sets. The use of filtering and neural network separation increases the accuracy of models trained on single-component audio signals, but it is not possible to fully compensate for the difference between data distributions. The model trained on multicomponent audio signals demonstrates extremely low accuracy when tested on single-component ones, which reveal the limitations of its direct application to the tasks of identifying individual sources. The results obtained emphasize the importance of forming realistic training sets and the need to develop hybrid approaches combining models for single-component and multicomponent signals in order to increase the versatility and stability of classifiers.

Results of a study of two methods for measuring the speed of objects with a magnetic field source based on the mutual secondary measurement conversion of signals from two induction sensors are presented. An example of such objects is a plasma piston in the channel of a magnetoplasmic electrodynamic accelerator, which is a moving conductor with a current that creates a magnetic field around itself. The first measurement method is based on an additive signal conversion function of two sensors, the second on a multiplicative one. A moving object is represented as a conductor with a current in the form of a linear rod of a certain length. The purpose of mutual conversion of two signals is to eliminate the dependence of the secondary signal on the coordinate of the object’s position. This makes it possible to ensure, provided the magnetic field of the object is constant (the current in the conductor is constant), that the secondary conversion function is strictly proportional to the measured parameter, the speed of the object. The methods of secondary measurement conversion were compared according to the criterion of ensuring the maximum length of the measuring section with a given permissible conversion error. The study was carried out by a computational experiment using computer modeling. Based on the study, it is concluded that the multiplicative method provides a longer length of the measuring section, compared with the additive method, with equal set values of the influencing quantities, and the parameters of functions approximating the experimental dependences of the length of the measuring section on the influencing quantities are obtained. It is assumed that the results will be used in further research on the development of methods for synthesizing devices for measuring the speed of objects with a magnetic field source.

Modern unmanned aerial vehicles can use various types of navigation systems to determine precise coordinates and orientation in space. An algorithm for visual navigation of an unmanned aerial vehicle based on a machine vision system and a pre-built map of the flight area is presented. A comparative analysis of approaches to navigation of unmanned aerial vehicles using computer vision is performed. Based on the results of the analysis, an approach is chosen to search for pairs of corresponding points on the terrain map and in the image from an unmanned aerial vehicle. Implementation options are considered, and based on the results of a comparative analysis, the ORB (Oriented FAST and Rotated BRIEF) algorithm with sorting of key point descriptors by Hamming distance is selected. Based on the results of field tests, an orthophotoplane is created. An algorithm is developed that outputs the latitude and longitude values with an error of up to two meters based on the compared image from an unmanned aerial vehicle with an orthophotoplane. The developed hardware and software complex can be used to determine an unmanned aerial vehicle location using a video camera.

A method for generating radar video frames in a two-position spatially distributed system of small-sized airborne radar stations is presented. Each station operates in the anterolateral viewing mode and applies a subaperture sliding window mode, in which the synthesized aperture is constructed from partially overlapping fragments. Such an organization makes it possible to ensure an acceptable frame rate without reducing their resolution. The achieved accuracy of determining the coordinates of objects on the sector boundary is almost four times higher than that provided by single-position survey methods. The results confirm the possibility of using the proposed method for all-weather operational control of zones in emergency situations.

The features of using a hydrostatic depth gauge to study the deformation and stability of profiles of mouth sections of small rivers are considered. Instruments and methods for determining the depth in river channels are analyzed. The method of performing experiments is described, as well as the main characteristics of the equipment used. The results of expeditionary research in the mouths of the Chernaya, Belbek, and Kacha rivers (Sevastopol) obtained using these technical means are described. The use of a hydrostatic meter allows you to quickly obtain information about the depth profile, which is necessary when calculating water flow and modeling the hydrological regime of river mouths.

A mathematical model of a gas-static support with active control is presented, designed to improve the accuracy of positioning and expand the functionality of measuring and diagnostic devices. The model includes a floating ring regulator, which allows for adaptive control of the support characteristics. The developed analytical model is validated by comparing it with the results obtained by the finite element method, which confirms its adequacy. The pressure distribution, static characteristics, and other parameters are calculated using the Delphi model. The static characteristics of the support, including load capacity and compliance in passive and adaptive modes of operation, are investigated. It is established that the use of active control can significantly increase the load capacity and reduce compliance. The modes of “zero” and negative compliance are identified, demonstrating high stability and positioning accuracy, as well as ensuring maximum load-bearing capacity under certain conditions. The factors influencing the stability of the support operation are considered, and the conditions for the occurrence of static instability when exceeding the control pressure limits are analyzed. The results obtained demonstrate the possibility of effectively controlling the characteristics of gasstatic supports by changing the control pressure, which opens up prospects for the use of these supports in precision measuring and diagnostic systems, as well as in other areas requiring high positioning accuracy and stability.

The problems associated with malfunctions in belt drives are considered, and a system for contactless monitoring of the drives condition based on ferromodulation sensitive sensors is proposed. The key factors of belt transmission failures have been investigated, including belt length changes, wear and slippage, fatigue deformations, and changes in the physical and mechanical characteristics of the belt material. The disadvantages of traditional monitoring methods are noted, in particular, the need to shut down measuring equipment, susceptibility to contamination, and insufficient information content of the data obtained. As a solution, a continuous and non-contact monitoring system is proposed based on the use of a complex of ferromodulation sensitive sensors: two sensors for monitoring belt movement, and two additional sensors installed in the area of the pulleys. This makes it possible not only to calculate the absolute belt elongation in a non-contact way, but also to diagnose critical phenomena such as belt slippage relative to pulleys and uneven loading. The proposed mathematical model makes it possible to calculate the instantaneous belt movement speed and its absolute elongation with high accuracy by analyzing the time delays between sensor actuations. The hardware of the system is described, including the device of sensor nodes and the principles of signal processing. Important advantages of the development are the ability to function in conditions of significant electromagnetic interference and variable loads, as well as continuous and contactless monitoring of belt drive malfunctions.

A model of a layered homogeneous medium in the form of “water–bottom soil” is considered. Taking into account different sets of types of bottom soil for the considered model, a dispersion equation describing the propagation of a longitudinal wave, as well as a dispersion equation for a transverse wave transformed from a longitudinal incident one, is obtained. Graphical dependences of the velocity of longitudinal and transverse waves on the thickness of the bottom soil layer are constructed. The influence of the emerging transverse wave on the overall wave process is estimated based on the analysis of reflection, transmission, and transformation coefficients.

The possibilities of deep learning for the detection and classification of nucleated cells of subvariants of acute lymphoblastic leukemia from digital images of bone marrow preparations without prior manual segmentation and feature extraction are investigated. During the study, a sample of 96 clinical cases is formed, including 9,468 images of nucleated cells from bone marrow preparations obtained using computer microscopy. The architecture of a single-stage neural network detector of the YOLO (You Only Look Once) family is used for the analysis, which significantly automated the process of cell detection and classification. The trained model demonstrates high overall accuracy — 0.98 in the test sample and 0.83 in the independent sample (48 clinical cases), which indicates a good generalizing ability and reliability of the presented method. The high sensitivity (1.0) for the B-cell subvariant and the specificity (0.83) for the T—cell emphasize the effectiveness and practical value of the proposed method in differential diagnosis, confirming its potential to improve diagnostic quality and accelerate data processing. Recommendations for use include the introduction of the developed model into scientific and clinical practice for further analysis and improvement in the course of research in the diagnosis of acute lymphoblastic leukemia.