论文部分内容阅读
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。
In order to study the multi-sensor fault diagnosis problem with unknown accurate mathematical model in aeroengine test, a fault detection method based on General Regression Neural Network (GRNN) is used to extract the constraint relationship between the sensors and the fault rule. A set of multiple-input-multiple-output GRNN is used to estimate the sensor output and generate residuals with the measured values. The suspicious sensors are compared with the threshold to find the GRNN with the smallest number of suspicious sensors in the neural network. The estimated signal of the suspicious sensor is used as a cross-validation of the reconstructed signal to other GRNNs. After verification, the suspicious sensor can be determined as the final fault sensor. In order to control the regression accuracy of the neural network, the multiple input multiple output neural network is decomposed into multiple multiple input single output networks. The feasibility of this method for sensor fault detection is verified by simulation data.