In this study, a deep learning based conditional density estimation technique known as conditional variational auto-encoder (CVAE) is used to fill gaps typically observed in particle image velocimetry (PIV) measurements in combustion systems. The proposed CVAE technique is trained using time resolved gappy PIV fields, typically observed in industrially relevant combustors. Stereo-PIV (SPIV) data from a swirl combustor with very a high vector yield is used to showcase the accuracy of the proposed CVAE technique. Various error metrics evaluated on the reconstructed velocity field in the gaps are presented from data sets corresponding to three sets of combustor operating conditions. In addition to accurate data reproduction, the proposed CVAE technique offers data compression by reducing the latent space dimension, enabling the efficient processing of large-scale PIV data.