关键词: 航空发动机；传感器；故障检测；故障隔离；广义回归神经网络

Abstract: The accurate mathematical model is not always clearly known and multiple sensors may be failure simultaneously in the field of aero engine experiments. In order to deal with this problem，a method of sensor fault detection based on a bank of general regression neural networks (GRNNs) is proposed. Firstly，a hypothesis of restriction relationship and regularity of sensor failures is extracted. Then，a bank of multi-input multi-output (MIMO) GRNNs are constructed and trained to compute the estimations of sensors. Next，the residuals of sensors are calculated with the estimations and measurements. The suspicious sensors are chosen by comparing the residuals and the thresholds. Moreover，the GRNN with the minimum number of suspicious sensors in the bank is found. Finally，the estimations are used as reconstructed signals in other GRNNs to cross verify suspicious sensors. The suspicious sensors are confirmed as fault sensors if the verification is passed. In order to better control regression precise，each MIMO GRNN is divided into multi MISO (multi-input single-output) GRNNs. The validity of this method is demonstrated with simulation experiments of single and multi fault sensors.

Key words: Aero engine；Sensor；Fault detection；Fault isolation；General regression neural network