Among the diverse Structural Health Monitoring (SHM) techniques, the Electromechanical Impedance (EMI) method using PZT sensors has been proven as a low-cost technique able to detect incipient damage. Additionally, it can be utilized as an actuator and sensor simultaneously. Because of its characteristics, it has emerged as a non-destructive approach to identify the incipient structural damage with greater sensitivity than the global SHM techniques. Data-driven algorithms based on EMI approach have been widely used in SHM applications for the health monitoring of civil structures. The combination of Machine Learning (ML) techniques with SHM techniques has shown an improved accuracy and reliability to identify and classify the damage from noise or misleading signals, such as EMI spectra. By means of supervised and unsupervised ML techniques and specifically Deep Learning (DL) approaches, anomalies of hidden patterns in large sets of data can be identified as a possible symptom of damage. However, one of the limitations of DL techniques is about the scarcity of real-world datasets and their frequent class imbalance distribution. This class imbalance issue can affect the predictive performance of machine learning models and a multiple representation of datasets involving different classes would be desirable. Therefore, a data augmentation technique should be applied to generate massive synthetic data, which will improve network performance. Different approaches have been proposed to augment and generate real-word datasets with the purpose of improving the robustness and performance of ML models. In this work, a Variational Autoencoder (VAE) model will be implemented to generate synthetic data corresponding to damage patterns, apparently not observed in an experimental EMI dataset, which will be used to feed a DL approach for localized damage identification in concrete structures using the EMI technique, which will provide the prior warning information from a PZT sensor. In the proposed approach, the artificial generation of synthetic data will be performed from a sampling of the latent space of the model VAE. This will allow covering a wide range of variations of the input data by leveraging probabilistic latent space and exploring all its dimensions. Data augmentation will make possible tackling limited and imbalanced real-world experimental datasets.