Cyber-physical systems (CPSs) represent a transformative paradigm that integrates physical processes with communication networks and computational technologies. These systems provide advanced functionalities that include remote monitoring and control, fault diagnosis, predictive maintenance, and automation. However, integration with cloud infrastructure introduces vulnerabilities to sophisticated cyber-attacks, alongside inherent system faults and disturbances, which complicates threat detection and classification. This paper proposes a hybrid machine learning approach for CPS, using the Energy Management System of a Connected and Automated Vehicle (CAEV) as a case study, with a focus on safety, reliability, and integrity. The first segment introduces an attack detection framework targeting false data injection attacks (FDIAs) using a combination of Duo-LSTM autoencoders and Light Gradient Boosting Machine (GBM). The second segment presents a fault detection approach that employs Bidirectional Long Short-Term Memory (Bi-LSTM)-based Recurrent Neural Networks to generate intermediate features and residuals, which are then used to train a Bi-LSTM classifier for identifying system faults. The framework triggers appropriate responses, such as risk mitigation for cyber-attacks and maintenance actions for system faults. Notably, the proposed cyber-attack detection framework achieved an accuracy of 99.89% in identifying FDIAs, while the fault detection framework demonstrated over 90% accuracy during simulations. These results highlight a proactive and effective approach to enhancing the security, reliability, and robustness of cyber-physical systems.