A data-augmented neural network approach is presented for classifying multipartite entanglement structures in continuous-variable (CV) quantum systems using homodyne measurement data. The key challenge addressed is generating sufficiently large and diverse training datasets for infinite-dimensional quantum systems with non-Gaussian states. A quantum data augmentation (QDA) method is proposed, leveraging physical principles such as entanglement invariance under mode permutation and convexity of separable states to expand training datasets without costly density-matrix recalculation. For tripartite states, classification accuracy improves from 0.961 to 0.986 with augmentation; for quadripartite states, from 0.796 to 0.928. The approach scales to five-mode systems (accuracy from 76.7% to 92.4%) and generalizes robustly to experimental noise and unseen state types.
