In mission-critical industries, early anomaly detection can make a difference between seamless operations and costly downtime. Traditional statistical methods often struggle with the complexity and high dimensionality of temporal dependencies, whereas Graph Neural Networks (GNNs) are emerging as a powerful tool for predictive maintenance, offering deeper insights into data from interconnected and distributed systems. Building upon our previously proposed detection model integrating Generative Adversarial Networks and GNNs, we improve and adapt it to a new semi-supervised setting. By transforming multivariate time series into static graph structures, our method separately captures spatial and temporal dependencies. Within an adversarial framework, we train an autoencoder on these graph representations to learn the system’s normal behavior. Anomalies are identified by analyzing the graph reconstruction errors and the discriminator scores. We test our approach through a real-life case study on industrial data from a gas turbine generator. A preliminary comparison with the baseline production model shows that our approach achieves higher precision, reducing both the frequency and duration of false alarms. Additionally, we conduct an ablation study by introducing synthetic anomalies through systematic connectivity alterations, assessing both the model’s robustness and the suitability of the graph representation algorithm. By emphasizing interpretability alongside detection accuracy, this study investigates the potential of GNNs for advancing predictive maintenance in complex industrial systems.