All issues

The work analyzes known classes of convolutional neural network models and studies selected from them promising models for detecting flying objects in images. Object detection here refers to the detection, localization in space and classification of flying objects. The work conducts a comprehensive study of selected promising convolutional neural network models in order to identify the most effective ones from them for creating mobile real-time computer vision systems. It is shown that the most suitable models for detecting flying objects in images, taking into account the formulated requirements for mobile real-time computer vision systems, are models of the YOLO family, and five models from this family should be considered: YOLOv4, YOLOv4-Tiny, YOLOv4-CSP, YOLOv7 and YOLOv7-Tiny. An appropriate dataset has been developed for training, validation and comprehensive research of these models. Each labeled image of the dataset includes from one to several flying objects of four classes: “bird”, “aircraft-type unmanned aerial vehicle”, “helicopter-type unmanned aerial vehicle”, and “unknown object” (objects in airspace not included in the first three classes). Research has shown that all convolutional neural network models exceed the specified threshold value by the speed of detecting objects in the image, however, only the YOLOv4-CSP and YOLOv7 models partially satisfy the requirements of the accuracy of detection of flying objects. It was shown that most difficult object class to detect is the “bird” class. At the same time, it was revealed that the most effective model is YOLOv7, the YOLOv4-CSP model is in second place. Both models are recommended for use as part of a mobile real-time computer vision system with condition of additional training of these models on increased number of images with objects of the “bird” class so that they satisfy the requirement for the accuracy of detecting flying objects of each four classes.

This paper investigates neural network models and polynomial models based on Chebyshev polynomials for interference compensation. It is shown that the neural network model provides compensation for parasitic interference without the need for parameter tuning, unlike the polynomial model, which requires the selection of optimal delays. The L-BFGS method is applied to both architectures, achieving a compensation level comparable to the LS solution for the polynomial model, with an NMSE result of −23.59 dB and requiring fewer than 2000 iterations, confirming its high efficiency. Additionally, due to the strong generalization ability of neural network architectures, the first-order method for neural networks demonstrates faster convergence compared to the polynomial model. In 20 000 iterations, the neural network model achieves a 0.44 dB improvement in compensation level compared to the polynomial model. In contrast, the polynomial model can only achieve high compensation levels with optimal first-order method parameter tuning, highlighting one of the key advantages of neural network models.

Use of algorithms based on neural networks can be inefficient for small amounts of experimental data. Authors consider a solution of this problem in the context of modelling of properties of ceramic composite materials reinforced with carbon nanotubes using perceptron complex. This approach allowed us to obtain a mathematical description of the object of study with a minimal amount of input data (the amount of necessary experimental samples decreased 2–3.3 times). Authors considered different versions of perceptron complex structures. They found that the most appropriate structure has perceptron complex with breakthrough of two input variables. The relative error was only 6%. The selected perceptron complex was shown to be effective for predicting the properties of ceramic composites. The relative errors for output components were 0.3%, 4.2%, 0.4%, 2.9%, and 11.8%.

This article studies a method of constructing a predictive neural network model of a time series based on determining the composition of input variables, constructing a training sample and training itself using the back propagation method. Traditional methods of constructing predictive models of the time series are: the autoregressive model, the moving average model or the autoregressive model — the moving average allows us to approximate the time series by a linear dependence of the current value of the output variable on a number of its previous values. Such a limitation as linearity of dependence leads to significant errors in forecasting.

Mining Technologies using neural network modeling make it possible to approximate the time series by a nonlinear dependence. Moreover, the process of constructing of a neural network model (determining the composition of input variables, the number of layers and the number of neurons in the layers, choosing the activation functions of neurons, determining the optimal values of the neuron link weights) allows us to obtain a predictive model in the form of an analytical nonlinear dependence.

The determination of the composition of input variables of neural network models is one of the key points in the construction of neural network models in various application areas that affect its adequacy. The composition of the input variables is traditionally selected from some physical considerations or by the selection method. In this work it is proposed to use the behavior of the autocorrelation and private autocorrelation functions for the task of determining the composition of the input variables of the predictive neural network model of the time series.

In this work is proposed a method for determining the composition of input variables of neural network models for stationary and non-stationary time series, based on the construction and analysis of autocorrelation functions. Based on the proposed method in the Python programming environment are developed an algorithm and a program, determining the composition of the input variables of the predictive neural network model — the perceptron, as well as building the model itself. The proposed method was experimentally tested using the example of constructing a predictive neural network model of a time series that reflects energy consumption in different regions of the United States, openly published by PJM Interconnection LLC (PJM) — a regional network organization in the United States. This time series is non-stationary and is characterized by the presence of both a trend and seasonality. Prediction of the next values of the time series based on previous values and the constructed neural network model showed high approximation accuracy, which proves the effectiveness of the proposed method.

This paper presents the results of experimental validation of some structural issues concerning the practical use of methods to overcome catastrophic forgetting of neural networks. A comparison of current effective methods like EWC (Elastic Weight Consolidation) and WVA (Weight Velocity Attenuation) is made and their advantages and disadvantages are considered. It is shown that EWC is better for tasks where full retention of learned skills is required on all the tasks in the training queue, while WVA is more suitable for sequential tasks with very limited computational resources, or when reuse of representations and acceleration of learning from task to task is required rather than exact retention of the skills. The attenuation of the WVA method must be applied to the optimization step, i. e. to the increments of neural network weights, rather than to the loss function gradient itself, and this is true for any gradient optimization method except the simplest stochastic gradient descent (SGD). The choice of the optimal weights attenuation function between the hyperbolic function and the exponent is considered. It is shown that hyperbolic attenuation is preferable because, despite comparable quality at optimal values of the hyperparameter of the WVA method, it is more robust to hyperparameter deviations from the optimal value (this hyperparameter in the WVA method provides a balance between preservation of old skills and learning a new skill). Empirical observations are presented that support the hypothesis that the optimal value of this hyperparameter does not depend on the number of tasks in the sequential learning queue. And, consequently, this hyperparameter can be picked up on a small number of tasks and used on longer sequences.

We present results of a study aimed at identification of a controlled object’s channels based on postprocessing of measurements with development of a model of a multiple-input controlled object and subsequent active modelling experiment. The controlled object model is developed using approximation of its behavior by a neural network model using trends obtained during a passive experiment in the mode of normal operation. Recurrent neural network containing feedback elements allows to simulate behavior of dynamic objects; input and feedback time delays allow to simulate behavior of inertial objects with pure delay. The model was taught using examples of the object’s operation with a control system and is presented by a dynamic neural network and a model of a regulator with a known regulation function. The neural network model simulates the system’s behavior and is used to conduct active computing experiments. Neural network model allows to obtain the controlled object’s response to an exploratory stimulus, including a periodic one. The obtained complex frequency response is used to evaluate parameters of the object’s transfer system using the least squares method. We present an example of identification of a channel of the simulated control system. The simulated object has two input ports and one output port and varying transport delays in transfer channels. One of the input ports serves as a controlling stimulus, the second is a controlled perturbation. The controlled output value changes as a result of control stimulus produced by the regulator operating according to the proportional-integral regulation law based on deviation of the controlled value from the task. The obtained parameters of the object’s channels’ transfer functions are close to the parameters of the input simulated object. The obtained normalized error of the reaction for a single step-wise stimulus of the control system model developed based on identification of the simulated control system doesn’t exceed 0.08. The considered objects pertain to the class of technological processes with continuous production. Such objects are characteristic of chemical, metallurgic, mine-mill, pulp and paper, and other industries.

Solution of Dynamic Light Scattering problem makes it possible to determine particle size distribution (PSD) from the spectrum of the intensity of scattered light. As a result of experiment, an intensity curve is obtained. The experimentally obtained spectrum of intensity is compared with the theoretically expected spectrum, which is the Lorentzian line. The main task is to determine on the basis of these data the relative concentrations of particles of each class presented in the solution. The article presents a method for constructing and using a neural network trained on synthetic data to determine PSD in a solution in the range of 1–500 nm. The neural network has a fully connected layer of 60 neurons with the RELU activation function at the output, a layer of 45 neurons and the same activation function, a dropout layer and 2 layers with 15 and 1 neurons (network output). The article describes how the network has been trained and tested on synthetic and experimental data. On the synthetic data, the standard deviation metric (rmse) gave a value of 1.3157 nm. Experimental data were obtained for particle sizes of 200 nm, 400 nm and a solution with representatives of both sizes. The results of the neural network and the classical linear methods are compared. The disadvantages of the classical methods are that it is difficult to determine the degree of regularization: too much regularization leads to the particle size distribution curves are much smoothed out, and weak regularization gives oscillating curves and low reliability of the results. The paper shows that the neural network gives a good prediction for particles with a large size. For small sizes, the prediction is worse, but the error quickly decreases as the particle size increases.

The paper provides a method for selecting the composition of a refrigerant with a given isobaric cooling curve using an artificial neural network (ANN). This method is based on the use of 1D layers of a convolutional neural network. To train the neural network, we applied a technological model of a simple heat exchanger in the UniSim design program, using the Peng – Robinson equation of state.We created synthetic database on isobaric boiling curves of refrigerants of different compositions using the technological model. To record the database, an algorithm was developed in the Python programming language, and information on isobaric boiling curves for 1 049 500 compositions was uploaded using the COM interface. The compositions have generated by Monte Carlo method. Designed architecture of ANN allows select composition of a mixed refrigerant by 101 points of boiling curve. ANN gives mole flows of mixed refrigerant by composition (methane, ethane, propane, nitrogen) on the output layer. For training ANN, we used method of cyclical learning rate. For results demonstration we selected MR composition by natural gas cooling curve with a minimum temperature drop of 3 К and a maximum temperature drop of no more than 10 К, which turn better than we predicted via UniSim SQP optimizer and better than predicted by $k$-nearest neighbors algorithm. A significant value of this article is the fact that an artificial neural network can be used to select the optimal composition of the refrigerant when analyzing the cooling curve of natural gas. This method can help engineers select the composition of the mixed refrigerant in real time, which will help reduce the energy consumption of natural gas liquefaction.

This paper proposes an architectural solution for organizing a convolutional neural network (CNN) oriented towards hardware implementation on edge devices under limited resources. To this goal, an approach to compressing spectrograms to a given size (28 × 28) is proposed using discretization, monoconversion, windowed Fourier transform, and two-dimensional interpolation. A balanced convolution procedure is developed based on compact convolutional filters, the size of which provides the balance between computational complexity and accuracy required for edge devices. An algorithm that enables convolution operations and calculation of the error function gradient in the convolutional layer in a single cycle ensuring increased performance in both inference and training modes of the CNN is proposed. The tradeoff between network trainability and its resistance to overfitting is optimized by applying the Dropout regularization method with a dropout coefficient of 0.5 for the fully connected layer.

The effectiveness of the proposed solution was demonstrated using the example of recognizing audio spectrograms of car and airplane engine sounds. The CNN was trained on a balanced dataset consisting of 7160 audio recordings. The trained network demonstrated high recognition accuracy (95%), low loss values (< 0.2), and balanced precision/recall/F-metric, demonstrating the effectiveness of the developed CNN model.

In this paper we consider the controlled motion of a helical body with three blades in an ideal fluid, which is executed by rotating three internal rotors. We set the problem of selecting control actions, which ensure the motion of the body near the predetermined trajectory. To determine controls that guarantee motion near the given curve, we propose methods based on the application of hybrid genetic algorithms (genetic algorithms with real encoding and with additional learning of the leader of the population by a gradient method) and artificial neural networks. The correctness of the operation of the proposed numerical methods is estimated using previously obtained differential equations, which define the law of changing the control actions for the predetermined trajectory.

In the approach based on hybrid genetic algorithms, the initial problem of minimizing the integral functional reduces to minimizing the function of many variables. The given time interval is broken up into small elements, on each of which the control actions are approximated by Lagrangian polynomials of order 2 and 3. When appropriately adjusted, the hybrid genetic algorithms reproduce a solution close to exact. However, the cost of calculation of 1 second of the physical process is about 300 seconds of processor time.

To increase the speed of calculation of control actions, we propose an algorithm based on artificial neural networks. As the input signal the neural network takes the components of the required displacement vector. The node values of the Lagrangian polynomials which approximately describe the control actions return as output signals . The neural network is taught by the well-known back-propagation method. The learning sample is generated using the approach based on hybrid genetic algorithms. The calculation of 1 second of the physical process by means of the neural network requires about 0.004 seconds of processor time, that is, 6 orders faster than the hybrid genetic algorithm. The control calculated by means of the artificial neural network differs from exact control. However, in spite of this difference, it ensures that the predetermined trajectory is followed exactly.