The application of clustering algorithms for resolving point clouds to reduce data size while retaining key geometric information is discussed. The algorithms such as DBSCAN, OPTICS, and k-means are analyzed. The focus is on evaluating the performance of algorithms for compressing point clouds containing millions of elements in terms of retaining object structure and optimizing computational resources. The results show that, despite the high accuracy of some algorithms such as OPTICS and DBSCAN, their performance is limited in terms of processing time and memory consumption, making them less suitable for large datasets. On the contrary, the G-Means algorithm has demonstrated high efficiency in resolving point clouds while retaining meaningful geometric features and providing acceptable processing speed. The proposed approach can significantly reduce the amount of data and maintain sufficient accuracy for subsequent modeling and analysis.The study was conducted using two examples: a power line support and an abandoned building. The G-Means algorithm reduced the point cloud file size by reducing the number of elements for both objects. The power line support point cloud originally had a more uniform structure, which allowed the rarefaction algorithm to remove more points with Object Geometry Retentions.The relevance of the problem of processing three dimensional unstructured data to create digital twins is due to the increased demand for efficient methods of designing and modeling of objects in various industries, such as construction, engineering, medical diagnostics, etc.
Automatic communication components visual recognition plays an important role in a number of modern applications, such as human-computer interaction, accessibility for individuals with hearing impairments, and advanced communication analysis systems. This technology enables detailed evaluation of verbal, non-verbal and paraverbal communication components. In particular, determining the number of phonemes based on articulation can provide systems with subtle paraverbal elements, thus improving both human and machine interpretation of communication. Previous developments in visual determining of minimal linguistic units in speech encountered significant challenges, particularly due to insufficient and imbalanced training datasets. For example, the disparity in data representation among classes was substantial, with the number of examples per class differing by factors of ten or more. Such an imbalance can cause bias in machine learning models, leading to skewed predictions and reduced overall accuracy, particularly for underrepresented classes. As a result, data balancing is essential for achieving more accurate and reliable results. In this context, the present study focuses on evaluating various data balancing techniques to address the task of visual speech recognition and counting the number of phonemes in English words. The considered model is based on ResNet-18 with 3D convolutions. The study examines the effects of both undersampling and oversampling approaches. Undersampling is performed by cutting both video fragments randomly in each class and video fragments of specific words entirely. Oversampling is carried out using classical video augmentation methods, specifically via VidAug library, and by cross-dataset augmentation. Experimental results show that these data balancing methods positively influence the training process, enhancing the model's ability to generalize across classes and improving its accuracy to a certain extent. However, they are not fully decisive in determining the model's overall effectiveness. The study highlights the need for further exploration of advanced balancing techniques and alternative model architectures to address the challenges posed by class imbalance comprehensively.
There are many tasks from various fields of science and technology that must process huge amounts of data in real time in current time. Reconfigurable computing systems based on programmable logic integrated circuits (FPGA) can become the technical basis for such computers. FPGAs are characterized by an increase in real performance due to an annual increase in the number of logic gates in new architectural solutions. At the same time, the increase in the number of contacts on the FPGA occurs disproportionately slowly compared to the growth of the volume of computing equipment on the chip. This leads to the fact that with the growth of the potential performance of the device, its switching and transmission capabilities remain at the same level. This paper examines an approach to increasing the throughput of data transmission channels in FPGAs by using lossless data compression algorithms multiple of the average compression ratio of the selected algorithm. It is proposed that an increase in the channel throughput leads to an increase in the specific performance of the reconfigurable computing system. A modernized Huffman algorithm is presented, which allows compressing dense data streams with an optimal compression ratio at the rate of arrival at the RCS. The use of new principles for constructing a Huffman tree and forming the lengths of new codes made it possible to reduce latency, make it equal to the duty cycle of data receipt and fit into the task cyclogram. A simulation was conducted, which showed that the developed algorithm makes it possible to compress a data stream at a speed of 128 Gbit/s on one computing core (pipeline). Comparison with other modifications of the Huffman algorithms on RCS showed that, with the same flow processing speed, the developed algorithm occupies the smallest hardware resource. This ensures a 25 % increase in specific performance. Also, the performance of this implementation exceeds all known third-party implementations on hybrid computing systems by up to 5 times. The use of such compression systems will increase the speed of information processing when solving such promising problems on RCS as training convolutional neural networks or pairing high-speed computers with external memory units.
Modern IT infrastructures are facing an increasing number of incidents, that require a fast and accurate response. In this regard, there is a growing interest in using large language models (LLM) to predict solutions automatically to eliminate IT incidents. A comparative analysis of four LLMs – ChatGPT, DeepSeek, Claude Sonnet, LLaMA – in the context of generating solutions based on textual descriptions of incidentsv is presented. A multi-criteria analysis with the participation of an expert group was used to assess the quality of the proposed models. The solutions were evaluated according to four criteria: accuracy, detail, practical applicability and conciseness. Based on 50 real-life incidents, responses were received from each model and then evaluated on a 5-point scale. The results showed that DeepSeek demonstrated the highest scores in accuracy and detail, making it the most suitable for complex technical incidents. Claude Sonnet and LLaMA distinguished themselves by their conciseness, providing quick perception of solutions in operational situations. ChatGPT showed good results in terms of applicability, but was inferior in terms of detail. A comparative analysis confirmed, that the choice of the optimal LLM depends on the specifics of the task: DeepSeek is preferable for deep diagnostics, whereas Claude Sonnet and LLaMA are suitable for rapid response. The data obtained can be used to improve the efficiency of IT incident management systems by choosing a more accurate model depending on the operating conditions.
Quantification of the interaction of technical facilities with natural environments requires the development of algorithms and methods of digital support. The development of CAD in the direction of forecasting the aerodynamic flow of architectural and construction objects makes it possible to simplify the task of determining wind loads. Even with the use of modern software systems, the initial data for determining wind loads is prepared individually for each facility. When analyzing high-rise through structures consisting of a large number of elements, the preparation of initial data for determining the static components of wind loads is a laborious and very demanding task. In the case of even small errors in determining wind loads, the results of the design calculation may have a significant error. Methods that simplify the collection of wind load on the lattice structure lead to a decrease in the accuracy of the calculation results. An option for automating the collection of the static component of the wind load on the tower is given in the article. Automating the collection of static wind load on a lattice tower reduces the likelihood of errors, labor costs, and calculation time for the bearing capacity of cellular lattice poles and power lines. The developed algorithm will be used to automate the collection of wind static load on equipment, elements of stairs, platforms and cable routes of cellular communication poles and power lines. The automated wind load collection algorithm will be useful in developing software for calculating lattice structures.
