Real-time prediction for the surge of turboshaft engine using multi-branch feature fusion neural network

The existing aeroengine instability precursor detection methods can be summarized as applying advanced signal processing technologies to various signals from the compressor test rig rather than the whole engine. Besides, these methods seriously depend on the artificial designed feature and threshold and also ignore the limit on the sensors onboard. Thus, with the help of the powerful feature extraction ability of the deep neural network, a real-time surge prediction method based on the multi-branch feature fusion neural network (MBFFNN) is proposed. First, the dataset can be obtained by using overlapping slices to divide surge test data into a sample sequence and using complete ensemble empirical mode decomposition with adaptive noise (CEEMDAN) to label each sample precisely. Second, for each sample, the time-domain statistical parameters are calculated and the recurrence plot is obtained by using phase space reconstruction. Finally, the MBFFNN with mixed data type input is designed, and its performance is evaluated by the generated dataset. The experimental results show that compared with multilayer perceptron (MLP), long short-term memory (LSTM), and deep residual network (DRN), MBFFNN has the best performance on two datasets for different surge tests, which demonstrates that the proposed method for surge prediction can accurately judge the state of the aeroengine, identify the instability precursor before the surge, and give an early warning in advance.