How Can AI Speed Up Access to LEO Satellites?

Low Earth orbit (LEO) satellites are reshaping internet, IoT and direct-to-device services. Yet, getting many devices to talk to a satellite is tricky: signals travel long distances, Doppler shifts are strong, and many users try to connect at once. Traditional random access methods struggle with these issues, causing delays and signal collisions.

A research team proposed a deep‑learning solution. They built a classifier that, using multi‑antenna data, instantly detects if several devices sent the same preamble (the start signal). The model, a lightweight 1‑D convolutional neural network, estimates how many users collided already in the early stage.

With this estimate, the system triggers an opportunistic transmission scheme: it adjusts each device's retry probability, optimizing resource use and cutting waiting times. Simulations following the 3GPP LEO standards show the new approach raises access success probability, lowers delay, better utilizes the shared uplink channel, and needs less computation than existing methods.

In short, AI becomes the "brain" that decides when and how devices talk to the satellite, making connections faster and more reliable for everyone.

The proposed framework leverages spatial correlation properties among the receiving antennas to extract discriminative features from the preamble signal. Practically, the signal captured by each antenna is represented as a time series; cross‑correlation between these series highlights patterns typical of multiple collisions. These features are fed into a lightweight 1‑D convolutional neural network (CNN) with few layers and small kernels, ensuring low computational overhead suitable for on‑board satellite hardware.

The CNN is trained on synthetic datasets generated with 3GPP LEO parameters, encompassing propagation delays (2‑30 ms), Doppler shifts (up to ±40 kHz), and simultaneous users (1‑100). Its output is a discrete estimate of the number of colliding users, which feeds an opportunistic scheduling algorithm based on a log‑linear utility function: \(U(p)=\log(1+\alpha p)\), where \(p\) is the transmission probability and \(\alpha\) an aggressiveness factor.

Balancing \(p\) and resource usage is optimized via a low‑complexity gradient descent algorithm executed in real time. Results show a 15‑20% increase in successful access probability compared to traditional back‑off schemes, and a 30 ms reduction in average delay. Moreover, the on‑board processor's energy consumption drops by 25% thanks to the lightweight CNN, making the solution scalable for large constellations.