Flight Attitude Prediction Model based on LSTM

Flight attitude is a key indicator of flight safety. In the event of a malfunction, the flight attitude of the aircraft may change rapidly. The traditional real-time feedback process of flight attitude does not have enough time for pilots to make decisions and perform actions. Using the advanced flight attitude prediction method, the pilot’s decision-making and decisionmaking time can be extended by several times, and the failure rate during the flight can be significantly reduced. In this study, an enhanced long short-term memory network architecture for flight attitude prediction is innovatively proposed. By constructing a two-layer structure with hierarchical spatiotemporal decoupling, the joint extraction and fusion of multi-scale dynamic features are realized. A random forest-driven multi-dimensional feature selection mechanism is introduced to effectively improve the association modeling ability of target parameters. A dynamic learning rate optimization strategy and a multi-branch prediction framework are designed to break through the bottleneck of time series expression of traditional models and significantly expand the early warning timeliness on the basis of maintaining the realtime prediction. Experiments show that the model can accurately predict the flight attitude, provide guidance for flight control, and improve flight safety.