A new type of signal’s representation that can be restored by artificial intelligence methods when parts of the data are lost

• Datum: 2026
• Hauptverfasser: Додонов, В. О., Гасанов, Е. І., Ізварін, Є. І., Отрох, С. І., Сегеда, І. В.
• Format: Artikel
• Sprache: Ukrainisch
• Veröffentlicht: Інститут проблем реєстрації інформації НАН України 2026
• Schlagworte: artificial intelligence; system survivability; neural networks; digital signal processing; data recovery; radio signal
• Online Zugang: https://drsp.ipri.kiev.ua/article/view/363162
• Назва журналу: Data Recording, Storage & Processing

The study develops a new type of radio signal’s representation adapted for recovery using artificial intelligence methods under conditions of partial data loss and destructive interference. It is demonstrated that traditional digital signal processing approaches based on classical I/Q representation do not provide sufficient survivability of information and computer systems operating in complex radio-frequency environments. This problem becomes especially critical under the influence of electronic warfare systems, multipath propagation, noise, and communication channel overloads, which lead to the loss of signal fragments and disruption of information flow integrity. The proposed approach is based on the transition from the classical time-domain complex representation of a signal to a structured time-frequency domain using the Fast Fourier Transform. The complex spectrum of the signal is decomposed into a logarithmic amplitude component and phase components represented through sine and cosine phase functions. This makes it possible to eliminate phase discontinuities and form a smooth feature space suitable for efficient processing by machine learning methods. Additionally, a spectral recovery mask is introduced to provide structural redundancy of the signal and enable reconstruction even in the case of partial data loss. The research investigates the application of convolutional neural networks, encoder-decoder architectures, and neuro-symbolic approaches for compensating damaged signal segments, eliminating phase imbalances, and suppressing noise. The models are trained on paired datasets containing distorted and reference signals, allowing the system to learn a generalized mapping between corrupted and reconstructed spectral representations. Unlike traditional interpolation methods, the proposed approach exploits the internal patterns and spectral structure of the signal, significantly improving reconstruction accuracy. The simulation results confirm that the new signal representation reduces the complexity of intelligent signal recovery tasks, increases processing speed, and provides a higher signal-to-noise ratio compared to classical digital signal processing algorithms. The proposed methodology improves situational awareness, resilience, and reliability of complex technical systems while creating a foundation for the development of next-generation self-recovering information systems capable of operating effectively under conditions of critical damage and unstable data transmission environments. Fig.: 3. Refs: 11 titles.