Digital Twin for Latency Prediction in Communication Networks
Authors
Melgaard, Magnus ; Taha, Adham ; N-A, Linette Susan Anil
Term
4. Semester
Education
Publication year
2023
Submitted on
2023-05-31
Pages
84
Abstract
Rapporten undersøger latens i trådløse netværk – den forsinkelse, der opstår mellem afsendelse og modtagelse af data – og afprøver en digital tvilling til at forudsige latens i realtid ved hjælp af neurale netværk. Der blev designet flere varianter af den digitale tvilling med forskellige input og forskellige antal modeller. For at indsamle latensmålinger blev der også udviklet en fysisk tvilling: en klient-server filoverførsel, der gav de nødvendige måleværdier. Den digitale tvilling blev trænet på oplysninger, der er kendt før en overførsel, såsom serverens fysiske placering og filstørrelse. Når de neurale netværk blev implementeret som TensorFlow Lite-modeller, kunne den digitale tvilling give forudsigelser på under et millisekund. Det var betydeligt hurtigere end den fysiske tvilling i alle scenarier, selv når den observerede latens var lavest. Forudsigelserne var vellykkede, og rapporten skitserer fremtidige tiltag for at øge nøjagtigheden, herunder at tage højde for tidsmæssige variationer i latensen.
This report examines latency in wireless networks—the delay between sending and receiving data—and explores a Digital Twin that predicts latency in real time using neural networks. Several versions of the Digital Twin were designed with different inputs and different numbers of models. To obtain latency measurements, a Physical Twin was developed: a client–server file transfer setup that provided the necessary data. The Digital Twin was trained on information available before a transmission, such as the server’s physical location and the file size. When the neural networks were deployed as TensorFlow Lite models, the Digital Twin produced predictions in under one millisecond. This was significantly faster than the Physical Twin in all scenarios, even when the measured latency was lowest. The latency predictions were successful, and the report outlines future steps to improve accuracy, including accounting for temporal characteristics in the observed latency.
[This abstract was generated with the help of AI]
Keywords
Documents