A rotating stall warning method for aero-engine compressor based on DeepESVDD-CNN

The compressor is one of the key components of the aircraft power system, and its rotating stall will greatly affect the performance of the whole system. Therefore, to ensure the stable and safe operation of the aircraft power system, it is necessary to develop a warning method for the compressor rotational stall. Based on the dynamic pressure signal collected from the compressor outlet pressure sensor of the aero-engine, this paper combines anomaly detection (AD) with deep learning to realize the compressor rotational stall warning system, which is named DeepESVDD-CNN. Specifically, a new anomaly detection method, deep ellipsoid support vector data description (DeepESVDD), is proposed to extract stall precursor based on dynamic pressure signals, and then to use the obtained warning signal as the input of convolution neural network (CNN) for training a classification model. This method only needs normal samples as training sets for unsupervised learning, and stores the characteristics of all normal samples in a deep network. The network parameters fully represent the stall warning model, which greatly simplifies the model training process and improves the performance of the stall warning model. The effectiveness and performance of this method are verified at four different compressor operating modes. In addition, two traditional warning methods for compressor stall are introduced and their performances are compared and tested. The experimental results show that DeepESVDD-CNN can identify the stall precursor and give the warning signal more accurately. It has good performance in the warning time, practicability, stability and real-time, and has good prospect in compressor stall warning tasks.