A.V. Liashenko. Models and Methods of Adaptive Routing in Transport IoT Networks. — Qualifying scientific work on manuscript rights.
Thesis for the degree of Doctor of Philosophy in specialty 172 — Telecommunications and Radio Engineering. — Educational and Scientific Institute of Telecommunication Systems of Igor Sikorsky KPI, Kyiv, 2026.
The dissertation addresses managing information flows of autonomous vehicles in a smart-city Internet of Vehicles (IoV) network, which unites mobile IoT nodes, roadside units, base stations and cloud services; a vehicle is a source of intense V2X messages, telemetry and sensor data, for which ETSI ITS and 3GPP C-V2X standards impose strict latency requirements.
The problem stems from spatio-temporal unevenness of load: during peak hours certain segments see a sharp rise in flow intensity while others stay moderately loaded; the load is non-stationary, noisy, and propagates across adjacent graph edges. Batch processing, sliding-window methods and EMA with fixed smoothing do not balance noise suppression and reaction speed; static path-search algorithms ignore the future network state, and topology-agnostic forecasting models fail to capture load propagation.
The problem is formalized as a time-dependent weighted IoV network Gₓ = (V, E, W(t), T), where vertices are infrastructure nodes (RSU, gNB, cloud aggregators) and the edge weight is the service time of the information flow on the segment, computed from measurable QoS indicators: the channel capacity coefficient η, the load coefficient L, and the load-change coefficient τ.
A telecommunication model of the IoV network and two interconnected methods were developed. The adaptive state-estimation method determines the smoothing coefficient α automatically from clustering of the QoS state (η, L, τ) using Mahalanobis distance, the Bhattacharyya coefficient and Takagi–Sugeno fuzzy inference, enabling reaction to load bursts without expert tuning. The forecasting and route-search method uses a graph neural network with Diffusion Convolution and GRU blocks as a time-dependent heuristic for A*, evaluating the expected QoS state of segments and reducing the search space while preserving route optimality.
The aim is to improve information-flow management efficiency for autonomous vehicles in an IoV network via the developed model and methods of adaptive estimation, forecasting and routing based on a modified A*. Seven tasks were solved: analysis of IoV architecture, study of limitations of existing methods, development of the model, adaptive-estimation method and forecasting method, improvement of route search, creation of a software complex, and experiments on the Irpin testbed and METR-LA.
The object of research is forming, transmitting and managing information flows of autonomous vehicles in an IoV network. The subject is methods of adaptive state estimation, forecasting and reactive-proactive routing of information flows.
The work applies graph theory, machine learning and neural networks, fuzzy set theory, simulation modelling, and software engineering. The implementation is a microservice system on Apache Kafka, an in-memory CSR graph, a TGNN forecasting service, and a real-time routing service.
Experiments confirmed the methods' effectiveness: the state-estimation error grows only 1.2 times under load bursts versus 2.7 times for the sliding window; on METR-LA reaction time is about 9 s versus 171 s, with +10% noise robustness at σ = 3 mph. The TGNN model has 205K parameters (32% fewer than DCRNN) with a 24% MAE gain (2.10 vs 2.77 mph). Route search cuts the search space by 56–68% with optimality loss under 0.00–0.08%. Under 583 msg/s (363,000 messages), the complex processes 96.6% of the flow at 1.19 ms latency versus 27.2% for full DB recalculation.
Scientific novelty: 1) for the first time, a telecommunication model of the IoV network as a time-dependent weighted directed graph was developed, with edge weight defined as the segment's flow service time based on measurable QoS indicators; 2) the adaptive state-estimation method was further developed, with automatic fuzzy rule-base generation from QoS clustering (Mahalanobis, Bhattacharyya, Takagi–Sugeno) without expert involvement; 3) the A*-based route-search method was improved using a time-dependent predictive heuristic, reducing the search space while preserving optimality and meeting ETSI ITS time constraints.
Practical significance: a software complex for IoV information-flow management was created, based on Apache Kafka, an in-memory CSR graph, adaptive-estimation services, TGNN forecasting and route search, suitable for high-frequency telemetry processing and real-time decision support.
Keywords: INTERNET OF THINGS (IOT), INTERNET OF VEHICLES (IOV), ADAPTIVE ROUTING, NEURAL NETWORKS, FUZZY LOGIC, TELECOMMUNICATION NETWORKS, INFORMATION FLOW MANAGEMENT, SIMULATION, SPATIO-TEMPORAL FORECASTING, SMART CITY.
