基于广义回归神经网络的传感器故障检测

Sensor Fault Detection Based on General Regression Neural Network

为了研究航空发动机试验中精确数学模型未知的多传感器故障诊断问题,采用基于广义回归神经网络(General Regression Neural Network,GRNN)组的故障检测方法,提炼出传感器之间的约束关系和故障规律,构建了一组多输入多输出GRNN,用于估计传感器输出,与测量值生成残差,通过与门限值比较判断可疑传感器,找到神经网络组中的具有最小可疑传感器数的GRNN。采用可疑传感器的估计信号做为重构信号交叉验证其它GRNN。通过验证即可确定可疑传感器为最终故障传感器。为了控制神经网络的回归精度,将多输入多输出神经网络分解为多个多输入单输出网络。通过仿真数据验证了该方法用于传感器故障检测的可行性。

The accurate mathematical model is not always clearly known and multiple sensors may be fail- ure 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 hypothe- sis of restriction relationship and regularity of sensor failures is extracted. Then, a bank of multi-input multi-out- put （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 suspi- cious 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 muhi MISO （multi-input single-output） GRNNs. The validity of this method is demonstrated with simulation experiments of single and multi fauh sensors.