Rotating machinery, integral to industries like manufacturing and power generation, requires reliable fault classification to prevent downtime and reduce maintenance costs. Traditional methods using vibration signal analysis face challenges due to the complexity of non-stationary signals and the scarcity of labeled data, often necessitating costly data collection. This paper presents a novel semi-supervised framework that leverages a Variational Autoencoder (VAE) to generate synthetic data, augmenting limited labeled datasets by capturing fault patterns in low-dimensional representations. Additionally, we employ the Generalized Synchrosqueezing Transform (GSST) to process vibration signals under variable speed conditions, enabling precise fault frequency detection without a tachometer. The synthetic data is integrated with a Convolutional Neural Network (CNN) for fault classification, enhancing model robustness. Experiments on the Case Western Reserve University (CWRU) bearing fault dataset demonstrate that our approach achieves 99.21% accuracy using only 20% labeled data. This label-efficient method improves fault diagnosis accuracy and generalizability, offering a scalable solution to reduce industrial maintenance costs and enhance operational efficiency through advanced deep learning techniques.