Autonomous anomaly detection is an increasingly critical requirement in the spacecraft (S/C) Attitude and Orbit Control Subsystem (AOCS). The current approach, based on hard-coded hierarchical thresholds, is effective for well-known failure cases but becomes inadequate when satellite dynamics grow more complex, often leading to exceptional cases that escape detection. Every missed anomaly forces the Ground Station into time-consuming investigations and the restoration of the satellite’s nominal operation. As an improvement over current practices, this work proposes a novel Machine Learning (ML)- based approach, specifically leveraging multivariate Recurrent Neural Networks (RNNs)—specifically, a Long Short-Term Memory (LSTM) network—to enhance the efficiency and reliability of onboard failure detection systems for AOCS. A key contribution of this study is the development of a synthetic multivariate dataset, designed using a flight-proven attitude simulator, which includes all necessary signals to capture the satellite’s complex attitude dynamics. This dataset encompasses multiple anomaly sources, accounting for inertia uncertainties, actuator deviations, flexible modes, and the effects of unpredictable exogenous phenomena and malfunctioning hardware—posing a significant challenge for AOCS anomaly detection systems. Finally, we provide a comprehensive taxonomy of anomalies, which has been used to ensure the coherence of the dataset design. Our experimental validation shows that our system promises to deliver better performance in terms of failure presence detection and faster identification with respect to the canonical approach.