This study aims to construct a high-precision, real-time UAV flight attitude stability prediction and evaluation system to solve the core problems of insufficient accuracy and poor real-time performance of traditional prediction methods. This is particularly significant for UAVs integrated with flexible textile sensors or those performing structural health monitoring of high-strength fiber-reinforced composite airframes. First, key attitude parameters such as pitch angle, roll angle, and yaw angle, as well as environmental data, are collected through multi-sensor fusion technology. After cleaning, feature extraction, and standardization preprocessing, a multi-layer LSTM prediction model combined with an attention mechanism is constructed. A multi-step rolling prediction strategy and a grid search-cross-validation hyperparameter optimization method are adopted. At the same time, a multi-dimensional stability evaluation framework covering indicators such as angular velocity variance, attitude deviation, and prediction error is established. Experimental results show that the proposed LSTM model significantly outperforms traditional ARIMA and other models in attitude prediction, reducing RMSE and MAE by 8.82% and 10.22%, respectively. It exhibits optimal prediction accuracy, particularly within the 25°-40°dive angle range, and effectively captures the temporal dependencies of attitude data. This study provides a scientific basis for optimizing UAV attitude control. Meanwhile, it holds significant theoretical and engineering value for improving fault early warning mechanisms.

Long Short-Term Memory, UAV, flight attitude, stability prediction, smart textiles