Fraud detection in financial transactions presents a persistent challenge due to extreme class imbalance and evolving attack patterns. While several machine learning (ML) and deep learning (DL) methods have shown promise, these solutions are fragmented and use traditional methods to address the severe class imbalance, leading to models with inflated metrics and poor generalization. In this study, we propose a unified ML-DL-XAI pipeline that integrates Variational Autoencoders (VAE) not only for data augmentation but also for feature engineering. Unlike traditional resampling, the VAE enables representation learning that preserves underlying data distributions while mitigating overfitting. Our pipeline incorporates interpretable machine learning models alongside neural networks to ensure both high performance and explainability. Empirical evaluations on a large-scale financial dataset demonstrate superior and reliable performance, that achieves an accuracy of 99.6%, a precision of 92%, and a recall of 85%, outperforming several recent benchmarks. By combining augmentation, feature engineering, and explainability in a single pipeline, this work offers a robust and practical answer for real-world fraud detection applications.