Deep Transfer Learning Based Time Series Data Prediction For Large Aircraft Test Flights

Aviation big data is one of the important research topics in the field of modern aviation industry.The acquisition of big data based on airborne sensors allows statistical analysis,support vector machines and artificial neural networks and other data mining and analysis technologies to be applied for abnormal detection of large aircraft.And state analysis provide more technical means.The time series data of large aircraft test flight is collected by sensors installed on each part of the aircraft.It has the characteristics of high dimensions,large number of samples,rapid changes in some data,and changes in sample distribution with flight conditions.Contains the information of the aircraft under various flight conditions,which has research value.Traditional prediction methods such as moving average and autoregression usually rely on only a small number of samples for modeling,and cannot fully extract the information assets contained in aviation big data,and most of the samples are abandoned.Deep learning methods have produced many excellent models for time series prediction problems,but they have strict requirements on input and output forms,and cannot be directly applied to time series data of large aircraft test flights.This thesis investigates the existing time series prediction methods,combines the characteristics of the data set,selects deep learning as the research direction,designs and implements a deep RNN-based time series prediction model,determines the content and form of input and output,and gives specific neural networks.Network model structure and hyperparameters.During the research process,in order to solve the problem that the samples in different dimensions of the data set are too large and the sample distribution changes with flight status,an additional data processing link is added in this thesis to group each dimension of the large aircraft test flight time series and divide them into different according to the flight status.The state data set simplifies model input and reduces prediction errors after model training.According to the characteristics of the hidden state of the RNN neurons in time series,this thesis also designs a cyclic output mechanism based on the network state,which significantly reduces the prediction error in the case of multi-step prediction.In order to further solve the problem that the sample distribution changes with flight status,this thesis conducts model generalization research based on deep transfer learning on two different flight status datasets,and applies deep transfer learning to the field of time series prediction for the first time.After investigating various transfer learning methods,this thesis designs a distribution loss function to participate in the training process of existing models,determines the calculation formula and data source of the distribution loss function,and analyzes the actual role of the function in model training.During the research process,in order to solve the problem of feature disappearance and model instability caused by the distribution loss function,this thesis specifically adds constraints to the function input,and proposes a regular optimized distribution loss function to make the model better on the target domain dataset Generalization capabilities.