This work presents a data reconstruction framework based on a modified variational autoencoder (VAE) architecture (Kingma and Welling 2022 (arXiv:1312.6114)) and compares its performance with a traditional B-spline (Chaudhuri 2021 (arXiv:2108.06617)) interpolation approach. The proposed VAE model, inspired by the architecture developed in Garola et al 2021 (IEEE Trans. Nucl. Sci.68 2165–72), incorporates two main enhancements: a stochastic masking layer that randomly removes 10% of the input vector components, and a β-annealing strategy to progressively reduce the influence of the Kullback–Leibler divergence term during training. The model has been trained on two datasets: (i) synthetic curves generated via Gaussian processes, and (ii) experimental profiles obtained from soft x-ray measurements in the RFX-mod reversed-field pinch experiment (Sonato et al 2023 22nd Symp. on Fusion Technology vol 66–68 pp 161–8). Both datasets consist of partially incomplete profiles with missing data points. The reconstruction performance of the VAE is evaluated and compared against B-spline interpolation. Results show that the VAE consistently outperforms the spline-based method, particularly under moderate to high levels of data loss, demonstrating the potential of deep generative models for robust data imputation in fusion plasma diagnostics.