Prediction of sensor data from flight tests using deep learning network and wavelet analysis

Abstract

Flight data is obtained through remote transmission and reception during flight tests, which can be subject to data acquisition failures due to communication disruptions or sensor malfunctions caused by the flight environment. Conducting a single flight test demands significant resources and personnel, while the acquired flight data holds considerable value, serving as a foundation for analyzing the vehicle's flying trajectory and meetness of the goals, refining future development directions, and addressing issues as they arise. During early development stages, numerous sensors are employed within the vehicle to maximize flight data acquisition. As development progresses toward completion, the number of measured sensors decreases as mission-related equipment is integrated. In instances where sensor measurements fail unexpectedly, or sensors are removed during development but their data is still required for further analysis, research has been conducted on methods to recover sensor data. Given the various frequency and time domain characteristics of sensor data, a wavelet-based analysis technique was employed to characterize the data and decompose sensor signals into multiple subbands. For each decomposed subband signal, a deep learning-based adversarial generation network, an attention-based long short-term memory structure of the encoder-decoder structure, and a dilated convolutional neural network were integrated to create a comprehensive prediction network. To compare the prediction performance of various deep learning network structures, data obtained from actual flight tests were utilized. Performance comparisons of the networks were conducted using both clean data and data with missing values, as flight data may occasionally contain missing values. Missing data rates ranged from 1.15\% to 17\%. Moreover, prediction results were analyzed in the time domain to compare each prediction characteristic, with the proposed WGLSTM framework consistently exhibiting superior performance. The versatility of wavelet analysis is demonstrated through the adaptability of various wavelet methods and the capacity to modify the mother wavelet to suit specific signal characteristics. Further analysis revealed that higher wavelet decomposition levels or modified wavelet functions could enhance prediction performance. Utilizing a complex wavelet function did not necessarily yield improved prediction performance. However, increasing the wavelet decomposition level while employing a complex mother wavelet function consistently outperformed predictions using the basic Haar wavelet function, as a complex wavelet function necessitates more wavelet decomposition to analyze and incorporate various characteristics into the prediction. In conclusion, this dissertation presents a framework for analyzing diverse sensor data with distinct characteristics and utilizing them to predict specific sensor data. The proposed method employs wavelet analysis and deep learning prediction techniques tailored to sensor characteristics and can be applied to supplement data during the development of a flying vehicle.