The flight control system, as the core safety assurance of a fighter jet, is responsible for controlling key actions such as flight attitude, heading, and altitude. Analyzing the monitoring parameters of the flight control system helps to detect anomalies in a timely manner, improving fault identification speed and ensuring flight safety. To address the challenges of traditional unidirectional neural networks in effectively capturing the temporal dependencies in complex flight control system data, and the insufficient robustness of the MSE loss function when handling anomalies and noise, this paper proposes a Bidirectional LSTM model, combined with Huber Loss to enhance noise resistance. For flight parameters, feature data alignment and down-sampling are performed first, and a sliding-window autoregressive prediction method is used to learn the normal flight patterns. The anomaly detection threshold is then set using Huber Loss to identify anomalies in the test set. Experiments were conducted on the ALFA dataset provided by Carnegie Mellon University, and the results show that the proposed improved Bi-LSTM model outperforms current state-of-the-art anomaly detection models across multiple metrics, particularly excelling in key indicators such as F1 score and AUC, effectively improving the model's anomaly detection capabilities.