【摘要】 为了提高无人机俯仰角故障数据处理和预测的精确性和可靠性，避免增加无人机试飞成本，利用长短期记忆网络(long short term memory, LSTM)、注意力机制+LSTM模型和差分自回归移动平均模型(autoregressive integrated moving average model, ARIMA)模型预测无人机试飞俯仰角故障数据。结果表明，ARIMA预测结果：平均绝对误差(mean absolute error, MAE)为0.35,均方根误差(root mean square error, RMSE)为0.73,平均绝对百分比误差(mean absolute percentage error, MAPE)为23.80%;LSTM模型预测结果：MAE=0.49,RMSE=0.74,MAPE=45.20%;注意力机制+LSTM模型预测结果：MAE=0.17,RMSE=0.53,MAPE=18.93%。可见注意力机制+LSTM模型比其余两种模型更适合于试飞俯仰角的数据预测，以上3种方法对无人机故障数据预测都具有实际意义，有效的预测可以推进自动飞行器和移动机器人的异常检测或外国直接投资研究的最新进展，以进一步提高自动和远程飞行操作的安全性。更多还原

【Abstract】 In order to improve the high accuracy and reliability of drone pitch angle fault data processing and prediction, and avoid increasing the cost of drone flight testing, long short term memory(LSTM), attention mechanism+LSTM model, and autoregressive integrated moving average model(ARIMA) model were used to predict pitch angle faults data in drone systems. The results show that for ARIMA, mean absolute error(MAE)is 0.35, root mean square error(RMSE) is 0.73, mean absolute percentage error(MAPE) is 23.80%. For LSTM model, MAE=0.49, RMSE=0.74, MAPE=45.20%. For attention mechanism+LSTM model, MAE=0.17, RMSE=0.53, MAPE=18.93%. It can be seen that the attention mechanism+LSTM model is more suitable for predicting the pitch angle of flight tests data than the other two models. The above three methods have practical significance for predicting unmanned aerial vehicle faults. Effective prediction can promote the latest progress in anomaly detection of autonomous aircraft and mobile robots or foreign direct investment research, in order to further improve the safety of automatic and remote flight operations.更多还原