91��ý

Every wireless communication system — from a smartphone call to a sensor in a smart city — depends on one deceptively simple task: extracting a meaningful signal from a background of noise. When conditions are ideal, this is manageable. When the signal is weak, the channel is crowded, or the computing hardware is limited, it becomes one of the hardest problems in modern radio engineering. A research project now underway at 91��ý Kazakh National University is tackling this problem directly, bringing neural networks and machine learning to bear on adaptive signal filtering and classification in radio systems. The project falls under the priority direction "Advanced Manufacturing, Digital and Space Technologies," in the specialised area of "Telecommunication Technologies and the Internet of Things." 

Signal classification in real radio environments is hard for a specific reason: the relationship between signal quality and classifier performance is highly nonlinear. Research by O'Shea, West, and others demonstrated that deep neural networks can classify signal modulation types with impressive accuracy — but primarily when the signal-to-noise ratio (SNR) is high. As SNR falls, accuracy degrades sharply. Existing architectures were not designed to adapt their behaviour in response to changing channel conditions. They learn a static mapping from signal to label, and that mapping breaks down when the environment shifts.

At the same time, the platforms on which signal processing must happen are increasingly constrained. IoT sensors, unmanned aerial systems, and edge devices in distributed networks cannot afford the compute budgets of a data centre. An algorithm that works beautifully on a high-performance server but fails on embedded hardware is not a deployable solution.

In Kazakhstan, systematic research into these combined challenges — low SNR robustness combined with cross-platform adaptability — has been limited. This project addresses both dimensions simultaneously.

The project's central scientific contribution is the development of adaptive algorithms: systems that do not merely apply a fixed learned function but dynamically adjust their parameters in response to real-time conditions. When the radio environment changes, the algorithm changes with it.

Two complementary technical streams underpin this ambition. The first is statistical signal filtering using distributional tests — specifically the Anderson-Darling test — and higher-order cumulants (HOC). These methods enable the algorithm to characterise the statistical structure of an incoming signal before attempting classification, providing a principled basis for adaptation. The second stream is neural network classification using convolutional neural networks (CNN) for feature extraction and recurrent neural networks (RNN) for sequence modelling. These architectures are trained on real-world signal data and are designed to distinguish modulation types while adjusting to changing channel conditions.

A defining design constraint of the project is cross-platform deployability. The algorithms are engineered to operate on high-performance computing infrastructure — multi-core Intel Xeon Gold 5218R processors paired with NVIDIA GPU accelerators — and on resource-constrained platforms including IoT devices and UAV onboard systems. This range is not a secondary consideration; it shapes every architectural decision in the project.

The project pursues four interconnected objectives. The first is the creation of a signal data library for training and testing models, drawing on signals collected from real radio systems with varying modulation types and noise profiles. This dataset forms the empirical foundation for everything that follows. The second is the development and implementation of adaptive filters for signal processing, based on distributional tests and HOC methods, optimised iteratively on the collected data. The third is the construction of neural network classification models — CNN and RNN architectures — trained and validated on real signal data, incorporating adaptive mechanisms that allow parameter adjustment in response to transmission conditions. The fourth is integration and real-world testing, in which the algorithms are embedded in real radio systems, their performance is evaluated, and the solutions are refined.

Primary data are collected using NI PXIe-1065, a multifunction platform for generating and digitally processing signals with various modulation types. This hardware enables data collection under realistic radio conditions and provides the experimental substrate for testing developed algorithms. MATLAB supports initial signal processing and modelling. TensorFlow and PyTorch provide the development environment for neural network architectures. High-performance servers with NVIDIA GPU accelerators handle the compute demands of model training on large datasets.

Result validity is maintained through statistical analysis, k-fold cross-validation, and independent holdout testing. All experimental procedures are documented to support reproducibility.

The project unfolds across three years. In 2025, the first quarter is dedicated to a comprehensive review of existing signal processing methods and identification of key gaps. From April to August, a baseline version of the filtering and classification algorithms is developed. From September to December, these baseline algorithms are tested on synthetic datasets covering multiple signal types and modulation schemes, generating the first set of quantitative performance benchmarks.

In 2026, the first half of the year is devoted to optimising the algorithms for low-SNR conditions and complex radio environments — the most technically demanding phase of the project. From July to October, the algorithms are ported and optimised for resource-constrained platforms, with particular attention to IoT deployment scenarios. November and December bring real-world system testing, validating performance under operational conditions.

In 2027, the final version of the adaptive algorithm suite is developed and refined for industrial deployment. From June to December, scientific papers are submitted to international journals, and the commercial and intellectual property dimensions of the results are formalised.

The tangible deliverable is a software package of adaptive algorithms for automatic signal processing, accompanied by technical documentation for integration into industrial radio systems and IoT devices. The package is designed to operate effectively across the full SNR range encountered in real deployments — not only under favourable conditions.

The practical applications are broad. Wireless communication systems operating at low SNR stand to benefit from more robust signal recognition. Military and civilian communication systems working under heavy interference load require reliable classification that conventional algorithms cannot deliver. IoT networks, where compute efficiency is as important as accuracy, need solutions that can run on the edge without sacrificing performance.

Scientifically, the project contributes to the intersection of signal processing theory, statistical methods, and deep learning — advancing knowledge of how neural architectures can be designed for robustness rather than merely optimised for peak accuracy. The results will be published in at least two articles in journals ranked in the top three quartiles by impact factor in Web of Science, or with a CiteScore percentile of at least 50 in Scopus, including target venues such as IEEE Transactions on Signal Processing and Signal Processing. The developed algorithms may also be licensed for commercial application by telecommunications companies and radio system manufacturers, opening a path from research to deployed technology.