An architectural approach to constructing metric spaces based on deep neural network representations for object identification in open-set environments

Context and Relevance. The paper proposes a hybrid architecture for time-series signal processing based on data representation as vector fingerprints in metric space and a fundamental separation between representation learning and decision-making processes. Unlike classical classification approaches, the system is not limited to a closed set of classes; instead, it implements the "Open Set Recognition" (OSR) paradigm, where unknown signals are treated as a normal environmental state rather than a classification error. Objective. To substantiate the effectiveness of a multi-level architectural approach to constructing metric latent spaces for object identification in open-set conditions. Hypothesis. The quality of latent space formation, its metric ordering, and identification reliability under uncertainty are positively correlated with the use of Radial Basis Function (RBF) structuring and negatively associated with traditional probabilistic classification methods (softmax) in tasks with incomplete prior knowledge. Methods and Materials. The study utilizes a dataset of time sequences reflecting the operation of fourteen different types of radio signal sources. The process of parameterization and temporal structure modeling is implemented using a hybrid neural network (CNN + LSTM) acting as a latent space encoder. The metric properties of the space were defined via an RBF layer with centroid initialization using the K-Means algorithm, while anomaly verification was performed through adaptive threshold control in Euclidean metric. Results. The results demonstrated that RBF-based structuring of the latent space enhances identification accuracy and improves the separability of embeddings from various signal sources. Furthermore, the density of embedding clusters correlates with the system's robustness to additive noise. The possibility of successfully rejecting unknown signals while maintaining high accuracy on known classes is shown. Conclusions. It is demonstrated that the hierarchical separation of feature extraction and geometric analysis in the latent space plays a critical role in the stability of identification systems. A metric approach is recommended for building scalable reference databases, facilitating flexible system adaptation to new sources without the need for complete retraining of neural network models

1. Воронцов, К. В. (2025). Метрические алгоритмы классификации. Лекции по машинному обучению. [Электронный ресурс]. URL: www.machinelearning.ru (дата обращения: 23.04.2026)
   Vorontsov, K. V. (2025). Metric classification algorithms. Lectures on machine learning. [Electronic resource]. URL: www.machinelearning.ru (viewed: 23.04.2026). (In Russ.)
2. Изотов Д.Ю., Паняев А.И. (2026а). Программа реализации радиально-базисного функционального слоя в TensorFlow/Keras для задач машинного обучения. Свидетельство о регистрации программы для ЭВМ RU 2026612861, 30.01.2026. URL: https://www.fips.ru (дата обращения: 23.03.2026).
   Izotov, D. Y., Panyaev, A. I. (2026). Software module for implementing a radial basis functional layer in TensorFlow/Keras for machine learning tasks. Software Registration Certificate RU 2026612861, 30.01.2026. URL: https://www.fips.ru (viewed: 23.03.2026). (In Russ.)
3. Изотов Д.Ю., Паняев А.И. (2026б). Программа инициализации центров радиально-базисного слоя искусственной нейронной сети. Свидетельство о регистрации программы для ЭВМ RU 2026612822, 30.01.2026. URL: https://www.fips.ru (дата обращения: 23.04.2026).
   Izotov, D. Y., Panyaev, A. I. (2026). Program for initializing the centers of the radial basis layer of an artificial neural network. Software Registration Certificate RU 2026612822, 30.01.2026. URL: https://www.fips.ru (viewed: 23.04.2026). (In Russ.)
4. Изотов Д.Ю., Паняев А.И. (2026в). Программа для адаптивного расчета параметров радиально-базисных функций (RBF) и центроидов классов в многомерном пространстве признаков. Свидетельство о регистрации программы для ЭВМ RU 2026660869, 15.04.2026. URL: https://www.fips.ru (дата обращения: 23.04.2026).
   Izotov, D. Y., Panyaev, A. I. (2026). A program for adaptive calculation of radial basis functions (RBF) parameters and class centroids in a multidimensional feature space. Software Registration Certificate RU 2026612861, 15.04.2026. URL: https://www.fips.ru (viewed: 23.04.2026). (InRuss.)
5. Карелина, М. Ю., Акулов, А. А., Акиньшин, Н. С., Подгорный А. В. (2025) Прогнозное моделирование в компьютерном зрении: гибридные архитектуры глубокого обучения дляанализа многомерных временных рядов. // Наукоемкие технологии в космических исследованиях Земли. – 2025. – Т. 17, № 6. – С. 51-61.
   Karelina, M. Yu., Akulov, A. A., Akinshin, N. S., Podgorny A. V. (2025) Predictive modeling in computer vision: hybrid deep learning architectures for the analysis of multidimensional time series. High Technologies in Earth Space Research, 17(6), pp. 51-61. (In Russ.)
6. Коротков В. Ф., Батчев С. А. Распознавание типа источника радиоизлучения с использованием метода радиальных базисных функций и теории приближенных множеств // Научная мысль. 2022. Т. 21. № 3-1 (45). С. 30–36.
   Korotkov, V. F., & Batchev, S. A. (2022). Identification of the type of radio source using radial basis functions method and the theory of approximate sets. Scientific Thought, 21(3-1), 30–36. (In Russ.)
7. Коротков В.Ф., Зырянов Р.С. Алгоритм распознавания типа источника радиоизлучения с использованием радиальных базисных функций. Изв. вузов России. Радиоэлектроника. 2017. - № 1. - С. 12-16.
   Korotkov, V. F., & Zyryanov, R. S. (2017). Source type recognition algorithm of radio emission using radial basis functions. Journal of the Russian Universities. Radioelectronics, (1), 12–16. (In Russ.)
8. Муаамар, М. А. M., Круглова, Л. В. (2026) Применение методов машинного обучения для классификации радиосигналов для управления беспилотными системами. // Ural Radio Engineering Journal. - 2026. - Vol. 10. № 1. - С. 23–45. https://elar.urfu.ru/handle/10995/149565
   Muaamar, M. A. M., Kruglova, L. V. (2026) Application of Machine Learning Methods for Radio Signal Classification in un-Manned Systems Control. Ural Radio Engineering Journal, 10 (1). pp. 76-81. https://elar.urfu.ru/handle/10995/149565 (In Russ.)
9. Рембовский, А. М. Радиомониторинг: задачи, методы, средства / А. М. Рембовский, А. В. Ашихмин, В. А. Козьмин. – 3-е изд., перераб. и доп. – Москва: Горячая линия - Телеком, 2012. – 624 с.
   M. Rembovsky, A. V. Ashikhmin, and V. A. Kozmin, Radio Monitoring: Problems, Methods and Equipment, 3rd ed. Moscow, Russia: Goryachaya Liniya - Telekom, 2012. (In Russ.)
10. Хайкин, С. (2008). Нейронные сети: полный курс, 2-е издание. М.: Вильямс, 1104 с.
    Haykin, S. (2008). Neural Networks: A Comprehensive Foundation, 2nd edition. Moscow: Williams, 1104 p. (In Russ.)
11. Amirian, M., Schwenker, F. (2020). Radial Basis Function Networks for Convolutional Neural Networks to Learn Similarity Distance Metric and Improve Interpretability. IEEE Access, 8, 123087–123097. https://doi.org/10.1109/ACCESS.2020.3007337
12. Bendale, A., Boult, T. E. (2016). Towards Open Set Deep Networks. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 1563–1572. https://doi.org/10.1109/CVPR.2016.173
13. Broomhead, D., Lowe, D. (1988). Radial basis functions, multi-variable functional interpolation and adaptive networks. Royal Signals and Radar Establishment Malvern (United Kingdom). https://www.researchgate.net/publication/233783084_Radial_basis_functions_multi-variable_functional_interpolation_and_adaptive_networks
14. Geng, C., Huang, S.-J., Chen, S. (2021). Recent Advances in Open Set Recognition: A Survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 43(10), 3614–3631. https://doi.org/10.1109/TPAMI.2020.2981604
15. Hassen, M., Chan, P. K. (2020). Learning a neural network classifier for open set recognition. Proceedings of the 2020 SIAM International Conference on Data Mining, 154–162. https://doi.org/10.1109/CVPR.2016.173
16. Oza, P., Patel, V. (2019). C2AE: Class Conditioned Auto-Encoder for Open-Set Recognition. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2302-2311. https://doi.org/10.1109/CVPR.2019.00241
17. Sainath, T. N., Vinyals, O., Senior, A., Sak, H. (2015). Convolutional, Long Short-Term Memory, fully connected Deep Neural Networks. IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 4580–4584. https://doi.org/10.1109/ICASSP.2015.7178838
18. Scheirer, W. J., Rocha, A., Sapkota, A., Boult, T. E. (2013). Toward Open Set Recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence, 35(7), 1757–1772. https://doi.org/10.1109/TPAMI.2012.256
19. Schroff, F., Kalenichenko, D., Philbin, J. (2015). FaceNet: A unified embedding for face recognition and clustering. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 815–823. https://doi.org/10.1109/CVPR.2015.7298682
20. Wang, Xiaoting. (2025). Automatic Modulation Recognition Based on Deep Learning. Journal of Computing and Electronic Information Management. 18. 76-81. https://doi.org/10.54097/834yjh71