Research On Forecasting Technology Of Flight Control Time Series Data Based On Temporal Convolutional Network

The prediction based on multi-dimensional time series data of flight control system can assist in prognostics and health management(PHM),realize the automatic assessment of the state of the aircraft,and predict the prognostics that may occur in the future through data monitoring and analysis,and realize the pre-maintenance of the aircraft,which is important to the maintenance of the aircraft.In this paper,a time-series data prediction technology based on time-convolution network algorithm is studied,and using this technology as the core algorithm,a set of flight control time series data prediction software system is written.The main contents are as follows.Firstly,the existing time series data processing algorithms are introduced.Recurrent Neural Network(RNN),Long-Short Term Memory(LSTM),Gate Recurrent Unit(GRU),and Convolutional Neural Network(CNN)are introduced,and the deficiencies of each algorithm used in the flight control time series data prediction task are analysised.Secondly,based on the Temporal Convolutional Network(TCN),a multi-dimensional Temporal Convolutional Network model is studied as the core algorithm for the prediction of flight control time series data.It has the characteristics of high accuracy and fast speed when processing complex and highly coupled time series data of flight control systems.The major improvements are as follows.Connect a one-dimensional TCN to form a multi-dimensional TCN network,so that the improved network model has the ability to process multi-dimensional time series data,and retain the network model's characteristics of causality and wildness.The activation function of the improved TCN model uses the parameterized ReLU function instead of the common ReLU function,which has the ability to differentiate the input less than 0.The improved TCN model uses a self-defined span connection method to connect the low-level convolution kernel and the high-level convolution kernel,so that the high-level convolution network can retain the details of the low-level network while obtaining a large receptive field.The improved TCN model is a full convolutional network,the last layer uses a convolutional layer to connect the convolutional layer and the output interface,so that the network can process any size of input data and get a specified format output.In addition,aimming at the characteristics of flight control time series data prediction task,the improved TCN algorithm and the existing timing processing algorithm,including RNN,LSTM and GRU,are verified and compared.The scheme design of each aspect is explained.The experimental results show that the improved multi-dimensional TCN network is significantly better than the existing processing algorithms of time series data such as RNN,LSTM,GRU in prediction accuracy and running speed.The effectiveness of the improved TCN algorithm used in the prediction of flight control time series data is verified.Finally,this paper proposes a set of software that can be used for the prediction of flight control timing data based on improved multi-dimensional TCN.A set of flight software system is written,including data selection and loading,algorithm selection,training parameter selection and adjustment,training and prediction function implementation,etc.The software integrates and embeds four existing algorithms that can be used for multidimensional time series data prediction,which are RNN,LSTM,GRU,and the multidimensional TCN algorithm studied in this paper,so that users can understand different algorithms and can choose algorithms flexibly according to actual needs.