Recurrent Neural Network Training For Large Signal Model Of TWT

Traveling wave tube(TWT)is a kind of microwave tube which is widely used.One of the core problems in the study of TWT is the beam-wave interaction of the TWT.Beam wave interaction theory includes small signal theory and large signal theory.The large signal theory is better to describe the nonlinear characteristics of the interaction process.Deep learning is a subset of machine learning,which obtains characteristics of data by training a dataset,and then classifies or predicts the data according to the characteristics.In this thesis,the deep learning neural network training method is applied to the beam wave interaction analysis of TWT.Several important and innovative results in this master dissertation are listed as below:1.A recurrent neural network model is built to train 8-18GHz single segment helix TWT and predict its output power.According to the similarity between the beam wave interaction of TWT and recurrent neural network,the neural network training model is established.Due to the strong nonlinear characteristics of the interaction process,introducing nonlinear features into neural networks by nonlinear transformation of TWT's input parameters and activation functions of neural network.The network is built with the deep learning computing model Tensorflow and the Python-based deep learning library Keras.The model is used to train 8-18GHz single segment helix TWT.The input of training data includes the frequency,working voltage,working current,coupling impedance,phase velocity and input power of TWT and the output power sequence varying along the tube length is used as the output.The trained model is used for prediction and good prediction results are obtained.2.Using recurrent neural network model to train 6-18GHz double helix TWT and predict its output power.After improving the training method of single segment helix,the data of double helix TWT with varying tube length are collected for neural network training.The output power of 6-18GHz double helix TWT with varying tube length can be well predicted by the training model.3.A TWT database is established,which can store the training data of recurrent neural network and parameters of TWT.The database stores operation parameters and output power of TWT,as well as TWT parameters such as high-frequency system and magnetic focusing system.Users can manage TWT data through a graphical interface program,including adding data,modifying data,inquiring data and others.