The use of unmanned aerial vehicles (UAVs) has been rapidly increasing in both professional and recreational settings, leading to concerns about the safety and security of people and facilities. One area of research that has emerged in response to this concern is the development of detection systems for UAVs. However, many existing systems have limitations, such as detection failures or false detection of other aerial objects, including birds. To address this issue, the development of a standard dataset that provides images of both drones and birds is essential for training accurate and effective detection models. In this context, we present a dataset consisting of images of drones and birds operating in various environments. This dataset will serve as a valuable resource for researchers and developers working on UAV detection and classification systems. The dataset was created using Roboflow software, which enabled us to efficiently edit and manipulate the images using AI-assisted bounding boxes, polygons, and instance segmentation. The software supports a wide range of input and output formats, making it easy to import and export the dataset in different machine learning frameworks. To ensure the highest possible accuracy, we manually segmented each image from edge to edge, providing the YOLO model with detailed and accurate information for training. The dataset includes both training and testing sets, allowing for the evaluation of model performance and accuracy. Our dataset offers several advantages over existing datasets, including the inclusion of both drones and birds, which are commonly misclassified by detection systems. Additionally, the images in our dataset were collected in diverse environments, providing a wide range of scenarios for model training and testing. The presented dataset provides a valuable resource for researchers and developers working on UAV detection and classification systems. The inclusion of both drones and birds, as well as the diverse range of environments and scenarios, makes this dataset a unique and essential tool for training accurate and effective models. We hope that this dataset will contribute to the advancement of UAV detection and classification systems, improving safety and security in both professional and recreational settings.

In today’s data-driven world, where information is one of the most valuable resources, forecasting the behavior of time series, collected by modern sensor networks and IoT systems, is crucial across various fields, including finance, climatology, and engineering. However, existing neural network models often struggle with time series forecasting collected by different sensors due to challenges such as large data volumes, long-term dependencies, noise, and anomalies, which can negatively impact predictive accuracy. This paper aims to enhance the accuracy of time series forecasting by proposing an adapted transformer architecture combined with an innovative data preprocessing method. The proposed preprocessing technique employs the fast Fourier transform (FFT) to transition from the time domain to the frequency domain, enriching the data with additional frequency-domain features. These features are represented as complex numbers, which improve the informational content of the data for subsequent analysis, thereby boosting forecasting performance. Furthermore, the paper introduces a modified transformer model specifically designed to address the identified challenges in time series prediction. The performance of the proposed model was evaluated using three diverse datasets collected by different sensors, each with varying measurement frequencies, data types, and application domains, providing a comprehensive comparison with state-of-the-art models such as LSTM, FFT-LSTM, DeepAR, Transformer, and FFT-Transformer. Extensive evaluation using five distinct performance metrics demonstrates that the proposed model consistently outperforms existing methods, achieving the highest accuracy across all datasets.

The importance of enhancing the accuracy of time-series forecasting using artificial intelligence tools is increasingly critical in light of the rapid advancements in modern technologies, particularly deep learning and neural networks. These approaches have already shown considerable advantages over traditional methods, especially due to their capacity to efficiently process large datasets and detect complex patterns. A crucial step in the forecasting process is the preprocessing of time-series data, which can greatly improve the training quality of neural networks and the precision of their predictions. This paper introduces a novel preprocessing technique that integrates information from both the time and frequency domains. To achieve this, the authors developed a feature extraction–extension scheme, where the extraction component focuses on obtaining the phase and amplitude of complex numbers through fast Fourier transform (FFT) and the extension component expands the time intervals by enriching them with the corresponding frequency characteristics of each individual time point. Building upon this preprocessing method, the FFT-LSTM forecasting model, which combines the strengths of FFT and Long Short-Term Memory (LSTM) recurrent neural networks, was enhanced. The simulation of the improved FFT-LSTM model was carried out on two time series with distinct characteristics. The results revealed a substantial improvement in forecasting accuracy compared to established methods in this domain, with about a 5% improvement in MAE and RMSE, thereby validating the effectiveness of the proposed approach for forecasting applications across various fields.

The modern development of the biomedical engineering area is accompanied by the availability of large volumes of data with a non-linear response surface. The effective analysis of such data requires the development of new, more productive machine learning methods. This paper proposes a cascade ensemble that combines the advantages of using a high-order Wiener polynomial and Stochastic Gradient Descent algorithm while eliminating their disadvantages to ensure a high accuracy of the approximation of such data with a satisfactory training time. The work presents flow charts of the learning algorithms and the application of the developed ensemble scheme, and all the steps are described in detail. The simulation was carried out based on a real-world dataset. Procedures for the proposed model tuning have been performed. The high accuracy of the approximation based on the developed ensemble scheme was established experimentally. The possibility of an implicit approximation by high orders of the Wiener polynomial with a slight increase in the number of its members is shown. It ensures a low training time for the proposed method during the analysis of large datasets, which provides the possibility of its practical use in the biomedical engineering area.

Accurate short-term forecasting of natural phenomena—such as temperature, electricity demand, and wind-power output—is critical for reliable planning in climate science and energy management. Forecasting accuracy can be enhanced by developing advanced feature engineering techniques, particularly through the use of wavelet transformation. This paper introduces a novel method to feature construction for short-term time series forecasting of natural phenomena, which is based on the use of the Stationary Wavelet Transform (SWT) and multi-family wavelets. The developed method allows for the extraction of signal characteristics while preserving the original data dimensionality, so each time series observation is supplemented with detailed coefficients from the wavelets. This approach includes the application of several wavelet families (e.g., Daubechies, Symlets, Coiflets, Haar, and Meyer), which increases the informativeness of the time series and improves the forecasting accuracy of neural network models, in particular LSTM. Experimental evaluation on three open datasets—meteorological variables, residential electricity demand, and wind-farm output. Relative to identically configured LSTMs trained on raw observations, the wavelet-augmented models cut error consistently: MAE by 13.6%, MSE by 17.7%, RMSE by 9.5%, and SMAPE by 13.9%. These improvements confirm that multi-family SWT features offer a dataset-agnostic route to higher short-term forecasting accuracy.

Time series forecasting is an important tool in many businesses. It can range from efficiently allocating resources for web traffic, predicting patient needs for staffing requirements, to forecasting a company's product sales. A particular use case, known as "cold start" forecasting, involves making predictions for time series that have little or no historical data, like a new product just entering the retail market. The key assumption of cold start forecasting is that products with similar characteristics should have similar time series trajectories. In such scenarios, traditional forecasting models that heavily rely on past observations may face challenges, necessitating the development of innovative approaches that can effectively make predictions in the absence of a substantial historical dataset. In this paper, Temporal fusion transformer neural network architecture was applied for solving cold start time series forecasting task. Modeling of the method was based on the use of a dataset contained in an open repository. After the preprocessing procedures, the dataset has about 370 time series, each of which has different length of series and has one categorical feature. Categorical feature have only 4 types of different values. For model training was performed to search for optimal hyperparameters across such parameters as: number of attention heads, learning rate, dropout percentage and hidden size.Model performed pretty well on this task. For model comparison were chosen metrics: MAE, RMSE, SMAPE. As can be seen from comparison with such popular models as DeepAR and LSTM, the proposed approach demonstrated the smallest forecasting error. Only one downside is that it can have more problems with anomalies in time series than DeepAR. But at the same time still provide interpretability of results. Прогнозування часових рядів є важливим інструментом у багатьох бізнесах. Воно може варіюватися від ефективного розподілу ресурсів для веб-трафіку, прогнозування потреб пацієнтів для персоналу, до прогнозування продажів продукції компанії. Особливий випадок використання, відомий як прогнозування "холодного старту", передбачає створення прогнозів для часових рядів, які мають мало або зовсім не мають історичних даних, як-от новий продукт, що лише виходить на роздрібний ринок. Ключове припущення прогнозування "холодного старту" полягає в тому, що товари з подібними характеристиками повинні мати схожі траєкторії часових рядів. У таких сценаріях традиційні моделі прогнозування, які сильно залежать від минулих спостережень, можуть зіткнутися з викликами, що вимагає розробки інноваційних підходів, які ефективно прогнозують у відсутності значного історичного набору даних. У цій статті була застосована архітектура нейронної мережі temporal fusion transformer для розв'язання завдання прогнозування часових рядів з "холодним стартом". Моделювання методу базувалося на використанні набору даних з відкритого репозиторію. Після процедур попередньої обробки, набір даних містив близько 370 часових рядів, кожен з яких мав різну довжину серії та одну категоріальну ознаку. Категоріальна ознака мала лише 4 типи різних значень. Для тренування моделі було проведено пошук оптимальних гіперпараметрів, таких як: кількість голів уваги, швидкість навчання, відсоток викидання та розмір прихованого шару. Модель показала досить хороші результати на цьому завданні. Для порівняння моделей були обрані такі метрики, як MAE, RMSE, SMAPE. Як видно з порівняння з такими популярними моделями, як DeepAR та LSTM, запропонований підхід продемонстрував найменшу помилку прогнозування. Єдиним недоліком є те, що він може мати більше проблем з аномаліями в часових рядах, ніж DeepAR. Але в той же час модель все ще забезпечує інтерпретованість результатів.