Convolutional Neural Networks (CNNs) are recognized as one of the most effective methods for accurate fault detection and classification in photovoltaic (PV) systems due to their ability to recognize complex patterns in image data. CNNs require large amounts of training data to perform optimally, which poses a challenge when dealing with insufficient data, particularly for less-common PV fault classes. This study employs Variational Autoencoders (VAEs) to generate synthetic data, significantly expanding smaller datasets. The focus was on underrepresented fault classes such as module open circuit and potential induced degradation, leading to over 450% increases in dataset size for these fault classes. The VAE-generated dataset was used to train CNNs within a fourteen-class classification framework. The results showed significant improvements in classification accuracy, validating the effectiveness of VAEs in addressing data scarcity and supporting CNNs in PV fault classification. This approach provides a scalable solution that maintains high standards in PV fault classification and analysis.