Improving Prediction Accuracy in LSTM Network Model for Aircraft Testing Flight Data

Abstract

Testing flight of new aircrafts is a critical step of the aircraft manufacturing process. The test flight will generate a huge amount of sensor data. Engineers can extract important information from these data and perform anomaly detection of the aircraft with the help of the information. The LSTM model currently in use during the test flight has many disadvantages. The main disadvantage is that the prediction result is not accurate enough, and the LSTM model's predicted value image does not fit the actual value image well. In addition, the input data dimension of the model is too high, resulting in a large amount of computation. However, the commonly used data dimension reduction methods will affect the accuracy of the results. Based on the above issues, the purpose of this paper is to optimize the LSTM network model for handling test flight data, and greatly improve the accuracy of the LSTM model prediction results. The optimization starts from the three perspectives of LSTM model parameters, LSTM model structure, and feature extraction in data preprocessing to improve the performance of the model.