These results highlight the potential of the Autoencoder–GCN pipeline as a scalable and reliable solution for AMS structure recognition under real-world constraints.