This study presents the use of the Transformer neural network to predict instantaneous voltage values in individual phases of a 220 kV power line. A sample of data was selected from measurements obtained from the Phasor Measurement Unit, containing 630,000 records (20 samples per second), representing a recording of 3 hours and 30 minutes. Despite the nonlinear nature of the data, the model achieved high accuracy – Root Mean Squared Error was less than 1%, with the absolute difference between the measured and predicted values ranging from approximately 10 V. The results suggest that the Transformer Neural Network architecture is capable of effectively modeling time dependencies even from relatively small sets of physical measurements, which opens up possibilities for further research in the field of predictive maintenance dynamics and stability control of power systems based on Phasor Measurement Unit data.

This article focuses on the design and implementation of a 3D Convolutional Neural Network (3D-CNN) for hyperspectral classification of tree species. Real data from repeated aerial imaging of selected sections of the Slovak electricity transmission system was used to test the models, using technology from our partner VUJE, a.s., which has a hyperspectral scanner consisting of two cameras, HySpex VNIR-1800 and HySpex SWIR-384, capturing wavelengths from 400 to 2 500 nm. The results of 3D-CNN classification based on autumn and spring data collection were compared, as well as classification after fusing data from selected areas for the purpose of comparing object vs. pixel classification. The presented object-based classification model based on 3D-CNN achieves on average 9 % better classification accuracy compared to pixel-based classification using a 1D Convolutional Neural Network (1D-CNN).

This paper aims to visualize hyperspectral data through effective dimensionality reduction using supervised autoencoders (SAE). With the Salinas dataset, several SAE architectures-both symmetric and asymmetric-are compared against Principal Component Analysis (PCA) as a reference approach. Initial experiments with linear SAEs confirmed their equivalence to PCA in visualization quality, though normalization improved both class separability and classification accuracy. Nonlinear activation functions (ReLU, sigmoid, tanh) were then explored, with sigmoid and tanh yielding more distinct class separation. The hybrid approach combining PCA preprocessing with the asymmetric SAE achieved over $99 \%$ classification accuracy and near-perfect separation, even for spectrally similar classes. The effectiveness of nonlinear asymmetric SAE architectures can be stated, especially when they are improved by PCA preprocessing.

The main goal of the paper was to present the importance of hyperspectral data visualization. Hyperspectral data are represented as objects in a multidimensional space. A data analyst working with such data has no intuitive understanding of their distribution or the topological proximity of individual classes and corresponding objects. The expected outcome of visualization is to enhance data comprehension for further processing, such as classification. In our study, we focused on dimensionality reduction and data visualization using Self-Organizing Maps. Class overlaps were identified, which was confirmed by evaluating the classification accuracy of the applied convolutional neural network.

This paper presents an approach to Coverless Image Steganography (CIS) leveraging Generative Adversarial Networks (GANs) for secure and imperceptible information hiding. Unlike traditional steganography methods that embed data into existing images, CIS methods generate new images directly conditioned on the hidden content, thereby avoiding detectable modifications. Using satellite imagery, we evaluate three encoder architectures, namely basic, residual, and dense, within the SteganoGAN framework to assess their effectiveness in preserving image quality and extraction fidelity. The results show the potential of GAN-based CIS and providing insights into the trade-offs between visual quality and hidden message capacity.

The main goal of this paper was to compare the hyperspectral classification results of tree species achieved with One-Dimensional Convolutional Neural Network, both with and without using Principal Component Analysis, using data from each camera (HySpex VNIR-1800 and HySpex SWIR-384) individually and then combined. The comparison showed that the use of Principal Component Analysis led to a substantial improvement in both classification accuracy and kappa score, highlighting the effectiveness of dimensionality reduction in improving model performance. If PCA is omitted, the classification results using the VNIR camera with 186 spectral channels are comparable to the results of the combined cameras with 474 spectral channels.

This paper deals with the investigation of the OneDimensional Convolutional Recurrent Autoencoder properties.The main objective of the presented experiments is to extract features from hyperspectral data and test on a simple classifier whether the resulting feature extraction has an effect on classification accuracy and kappa score.By adding autoencoder layers,the goal was to analyze to what extent it is possible to reduce the dimensionality of the data in a way that does not affect the classification results.Salinas,Indian Pines and Pavia Centre hyperspectral remote sensing scenes were used for the purpose.The datasets were tested without dimensionality reduction and then after dimensionality reduction,to minimize the reconstruction error.The results showed that by testing across multiple benchmark datasets with a simple classifier the chosen autoencoder was able to reduce the dimensionality of all the datasets listed up to an $8: 1$ compression ratio while maintaining classification performance,highlighting its effectiveness and generalizability.

This paper is focused on the comparison of two feature space dimensionality reduction approaches, namely Principal Component Analysis and One-Dimensional Convolutional Neural Network to identify the hyperspectral classification model with the highest classification accuracy. The reference model is an enhanced multilayer perceptron without using dimension reduction. It differs from the standard multilayer perceptron by adding dropout layers, which prevent overfitting and stabilize the training process. The main idea of choosing a One-Dimensional Convolutional Neural Network approach was to leverage its capability to reduce the feature space of the input data using a form of supervised learning. This way, it was possible to extract class-relevant features of the input data at a pixel level. The results show that the kappa score and the accuracy of the proposed model were 0.828 and 0.803 without dimension reduction, 0.863 and 0.843 with Principal Component Analysis, and for One-Dimensional Convolutional Neural Network 0.847 and 0.825, respectively.

This paper presents a federated learning approach to the classification of Distributed Denial of Service (DDoS) attacks using convolutional neural networks. The study employs the CIC-DDoS2019 dataset, which includes a variety of benign and malicious network traffic data. The classification task focuses on identifying specific types of DDoS attacks, such as Domain Name System (DNS), PORTMAP, and SYN Flood, alongside non-DDoS traffic. Utilizing MATLAB’s parallel server architecture, the research implements federated learning to train local models on distributed datasets, enhancing data privacy and security. The federated learning model achieves a global accuracy of 0.894 after 58 communication rounds (174 epochs), compared to a standalone training accuracy of 0.9607 after 300 epochs. While standalone training benefits from centralized data leading to higher accuracy, federated learning offers significant advantages in terms of scalability, privacy preservation, and utilization of decentralized computational resources.

This paper proposes a YOLO-based detector designed to identify warning signals preceding hazardous conditions on high-voltage transmission lines, especially those that degrade and damage insulators on the power lines. Acoustic emissions recordings are used for this purpose, which were obtained by the monitoring system we have designed, built and installed for this purpose. The identification of partial discharge in the acoustic emissions of the transmission system is one of the opportunities to prevent the occurrence of undesirable consequences. The primary goal is to create an auxiliary tool to help identify acoustic recordings that potentially contain waveforms representing the occurrence of partial discharge on an insulator. The created detector works on the basis of a neural network trained on a set of partial discharge synthetic data mixed with real-world on-site noise recordings, transformed and labeled by a dataset generator.

This work deals with the implementation of Bidirectional Long Short-Term Memory and Gated Recurrent Unit neural networks to detect anomalies in the form of DDoS attacks in the CIC-DDoS2019 dataset. The proposed approach is based on the regression models of the mentioned neural networks, which generate a certain reconstruction error for the samples fed into their input. The intention was to achieve a smaller reconstruction error for the network flow samples not containing a DDoS attack compared to samples containing such an attack. The reconstruction error is evaluated using a threshold to decide whether an attack is present. The result of this work is a comparison of two validated recurrent neural network models in network traffic anomaly detection for the CIC-DDoS2019 dataset. The comparison was performed using commonly used neural network evaluation metrics.

This paper focuses on the design and implementation of a one-dimensional convolutional neural network model in hyperspectral tree species data classification. The proposed approach involves dimensionality reduction in the form of feature extraction using principal component analysis. Both the number of components, which influences the dimensionality of feature space and explained variance, which indicates the proportion of data variability captured were selected as the criterion for feature extraction. As a result, this work compares the classification accuracy and kappa scores of the proposed model for data without using principal component analysis and for data using principal component analysis with different number of transformed input features depending on the criterion used.

Hyperspectral imaging has become an active area of research and development in the world. In the past, hyperspectral imaging was only available to selected research teams in specific areas such as space research. Nowadays, the use of hyperspectral imaging systems is increasingly entering commercial use. Given the current technical capabilities, hyperspectral imaging can be implemented in the performance of aerial surveys of outdoor power lines

This paper introduces a model for the authentication of large-scale images. The crucial element of the proposed model is the optimized Pulse Coupled Neural Network. This neural network generates position matrices based on which the embedding of authentication data into cover images is applied. Emphasis is placed on the minimalization of the stego image entropy change. Stego image entropy is consequently compared with the reference entropy of the cover image. The security of the suggested solution is granted by the neural network weights initialized with a steganographic key and by the encryption of accompanying steganographic data using the AES-256 algorithm. The integrity of the images is verified through the SHA-256 hash function. The integration of the accompanying and authentication data directly into the stego image and the authentication of the large images are the main contributions of the work.

Reliability, longevity, and maintenance costs of electric power industry installations and equipment depend strongly on the extent to which their design reflects relevant environmental factors, such as expected levels of local environmental pollution. These factors guide the choice of specific types of components – insulators, towers, conductors, etc. – and are often estimated through complex and tedious long-term field measurements of pollution deposits. In Slovakia, such field measurements were mandated by the national standard STN 33 0405. This standard was retired in 2015 without replacement, which opened the way for developing alternative and less cumbersome methods. One such alternative is to apply artificial intelligence techniques to atmospheric pollution and other relevant data, which is already routinely monitored and collected in many countries. In this paper, we explore the strength of the relationships between the field measurements performed in various regions of Slovakia according to STN 33 0405 and atmospheric pollution data monitored and collected by the Slovak Hydrometeorological Institute (SHMÚ). The paper is focused on input attributes significance, in relation to output attributes. It represents the first phase of our long-term research aiming at the creation of reliable regression models of local pollution in order to replace the cumbersome field measurements mandated by STN 33 0405.

This paper addresses the regression modeling of local environmental pollution levels for electric power industry needs, which is fundamental for the proper design and maintenance of high-voltage transmission lines and insulators in order to prevent various hazards, such as accidental flashovers due to pollution and the resultant power outages. The primary goal of our study was to increase the precision of regression models for this application area by exploiting additional input attributes extracted from satellite imagery and adjusting the modeling methodology. Given that thousands of different attributes can be extracted from satellite images, of which only a few are likely to contain useful information, we also explored suitable feature selection procedures. We show that a suitable combination of attribute selection methods (relief, FSRF-Test, and forward selection), regression models (random forest models and M5P regression trees), and modeling methodology (estimating field-measured values of target variables rather than their upper bounds) can significantly increase the total modeling accuracy, measured by the correlation between the estimated and the true values of target variables. Specifically, the accuracies of our regression models dramatically rose from 0.12–0.23 to 0.40–0.64, while their relative absolute errors were conversely reduced (e.g., from 1.04 to 0.764 for the best model).

Presented work summarizes selected Convolutional Neural Networks classification of interfering elements in the meteorological images. Interfering elements, such as raindrops and insect adhered to camera lens, bright sun and other elements limit the process of automatic remote estimation of visibility at airports. We have experimented with three groups of pretrained neural networks. Namely we used AlexNet, DenseNet and ResNet. DenseNet169 classification appears to be a suitable solution. All the examined classification metrics, under the conditions of a classification threshold of 99% and above, indicated values above 90%. The paper also presents real deployment of classification models for full high definition camera images.

The paper is focused on the proposed model of image authentication. The model is based on the steganography principle using a neural network, symmetric encryption and cryptographic hash functions. A key element of the software module is the Optimized Pulse-Coupled Neural Network Model (OM-PCNN). The neural network generates position matrices for embedding authentication data into the cover images with an emphasis on image entropy. In order to increase the security of the proposed solution, the neural network weights are initialized using a steganographic key and, in addition, the authentication data is encrypted by the AES-256 algorithm. Image integrity is tested using SHA-2 hash function with 512-bit hash.

The paper aims to the evaluation of image quality assessments of stego images based on entropy. Two embedding approaches are compared. The first approach is based on a position matrix, which is generated for each image using the Optimized Model of Pulse Coupled Neural Network (OM-PCNN). The second, so called reference approach, is based on generating the random positions for embedding. The subject of research was to observe the increase in entropy of stego images compared to cover images for both embedding approaches. From the point of view of image steganography, a case with zero change in entropy is considered an ideal result. Experiments have shown that the embedding by OM-PCNN position matrix causes smaller increase in entropy compared to the random embedding. Therefore, the OM-PCNN approach is prerequisite for the lower detectability of the message embedding.

This paper presents an implementation of Convolutional Neural Network (CNN) symmetric encryption. The results based on selected criteria are compared with Advanced Encryption Standard (AES). Authors are testing both implementations using the set of differently sized and content independent files. Compared metrics include the execution time, memory occupancy and processor load. Robustness of CNN encryption against the crypto attacks was not tested and it is out of the scope of this paper. Concluded results offer the usability prospects of selected neural network for symmetric cryptographic purposes. The results of the experiments have shown that CNN has the potential for cryptographic purposes under the conditions specified in this paper.