A style transfer augmentation method for liquid rocket engine turbopump diagnostic data

Focus on the problem that fault samples of liquid rocket engine turbopump are few and difficult to obtain， this paper proposed a style transfer and augmentation method for turbine pump fault simulation data based on one-dimensional convolutional neural networks （1D-CNN）. A 1D-CNN model with multi-layer convolutional structure was constructed， which was pre-trained by using the reconstructed input object information optimization method. The measured normal data of the turbine pump is used as the style information input for the 1D-CNN， while the simulation fault data of the turbine pump is used as the content information input for the 1D-CNN. The hidden layer information of the two inputs is then obtained respectively.Using the randomly generated white noise signal as the initial value of the 1D-CNN training-generated signal， the style loss and content loss functions were constructed. Through multiple error backpropagation calculations of the two loss functions， the synthetic fault data that simulated the actual emission conditions of turbine pump was generated. The experimental verification of the turbine pump bearing fault test of the liquid rocket engine indicates that the synthetic signals of the bearing rollers and the outer ring faults generated by this method had cosine similarity degrees of 0.703 and 0.62 respectively with the experimental signals. The coherence of the fault characteristic frequencies was close to 1， which could meet the actual application requirements and replace the measured fault signals for the fault diagnosis of the turbine pump device.