Data Augmentation and Recognition for Plunger Lift Based on Variational Autoencoders

Plunger lift technology effectively addresses liquid loading issues in the later stages of natural gas well production. With the continuous development of artificial intelligence technology, the rapid identification of plunger data has become a challenge in field operations. Due to the lack of labeled field data and significant data imbalance, it is difficult to directly identify using discriminative models. This paper proposes a plunger lift generative model based on variational autoencoders (VAEs) for data augmentation of pressure data in plunger lift processes. The model maps original data through an encoder to obtain a high-dimensional latent variable probability distribution; the decoder maps the latent variable back to the real space to generate samples. The generated data supplements the original data, which is then input into a discriminative network for training. The results show that compared to the original discriminative model, the model’s performance indicators have improved after data augmentation.