Cyber-physical Systems (CPS) such as power grids, transportation, water supply, and health systems are often targeted by cyber-attacks due to their increasing automation and interconnectivity. Ensuring cybersecurity for these CPS is essential as compromising them could disrupt vital services with potentially extensive consequences. Many studies on cyber-attack detection in CPS focus on domain-specific solutions, which limit their applicability across different CPS environments. Addressing this challenge requires robust cross-domain anomaly detection methods that can generalize across multiple CPS domains. This study presents a hybrid approach for cross-domain anomaly detection in CPS using a Domain-Adversarial Neural Network (DANN) enhanced with an autoencoder and Mutual Information Maximizer (MIM) to improve domain-invariant feature learning. The proposed methodology leverages four distinct CPS datasets: water distribution testbed, gas pipeline, power system and Hardware in the Loop Industrial Control System (HAI) datasets. A comprehensive hyperparameter tuning process is conducted to optimize the architecture, with multiple configurations evaluated based on classification accuracy and F1-scores across domains. The model performance is assessed on gas pipeline, power system and HAI datasets, and the best configuration is selected for anomaly detection in a cross-domain CPS context. Results highlight the model’s potential to transfer learned representations effectively across different CPS domains.