Using Machine Learning Tools for Rotating Stall Warning in a Contra-Rotating Compressor

Abstract

This paper takes a low-speed axial contra-rotating compressor as the experimental object, and the sensor array is used to collect the pressure sequences in stall conditions for different speed configurations. These pressure data sets are then preprocessed to train the neural networks. A self-learning stall threshold method based on kernel density estimation (KDE) is utilized to obtain the alarm thresholds. By utilizing the best-performing long short-term memory (LSTM) model to predict the stall initiation time for 15 speed configurations with different stall characteristics, the results show that the model can provide early warning before stall for 11 speed configurations. For the rest four speed configurations, the stall initiation time predicted by LSTM is unsatisfactory. To overcome the poor generalizability of LSTM, a convolutional neural network (CNN) combined with LSTM (CNN–LSTM) stall warning method is developed. The stall warning results indicate that the CNN–LSTM has a better capability in fitting the nonlinear pressure stall data and issues warnings before a stall occurs for all speed configurations. By comparing the pressure time series predicted by LSTM and CNN–LSTM, it is obvious that the CNN–LSTM is more sensitive to perturbations than before stall occurs.