The Grand Radio Array for Neutrino Detection (GRAND) is an envisioned observatory that aims to detect the radio emission from air showers induced by ultra-high-energy cosmic particles; in particular, by neutrinos. Because these are rare, GRAND requires a large detection area, necessitating the use of inexpensive radio-detection units that must trigger autonomously. Such a trigger must achieve a high rejection efficiency of the dominant transient radio background, while keeping a high detection efficiency for air shower radio pulses. Fortunately, air shower simulations and field data suggest that air shower radio pulses exhibit characteristic features whose exploitation would lead to a powerful background rejection. We present the results of a machine learning signal classification method that has been tested on simulations and data recorded by a GRAND prototype in the Gansu province of China. Considering time traces that pass a simple $3\sigma$-transients pre-trigger, a neural network is able to keep 66% of the air showers pulses for a Signal to Noise Ratio (SNR) between 3 and 4, and more than 86% after a SNR of 4, while rejecting between 97% and 99% of the background traces. This trigger method will eventually be implemented to the next prototypes to be tested under field conditions.