Thrust Drop Fault Diagnosis in Launch Vehicles Using Model–Data Approach

This study investigates fault detection and diagnosis for launch vehicles experiencing thrust drop faults (TDFs). A model–data hybrid approach, integrating model-based estimation and data-driven learning, is proposed for real-time estimation of thrust magnitude and faulty engine location. Given that thrust magnitude primarily influences translational dynamics, while faulty engine location affects attitude motion, FDD is divided into two subtasks: estimation of the thrust magnitude and location of the faulty engine. Considering that both thrust and mass are unknown and coupled together, a joint estimation algorithm integrating Kalman filtering with recursive least squares is proposed, enabling the concurrent estimation of thrust magnitude and mass. Additionally, a location diagnosis strategy based on the temporal convolutional network (TCN) is proposed. To improve the correct rate of faulty engine location diagnosis, we differ from traditional methods using apparent acceleration and attitude angle deviation signals. Furthermore, we use thrust magnitude and fault occurrence time obtained from the model-based method for training the TCN model, improving the correct rate. Simulation results demonstrate the effectiveness of the proposed method, achieving thrust estimation error below 2%, fault localization accuracy of 99.03%, mass estimation error under 0.0025%, and low-latency diagnosis. This approach offers a solution for in-flight TDF diagnosis under limited measurable information and parameter uncertainty.