基于神经网络集成的传感器故障诊断研究

针对诊断传感器偏置故障及漂移故障的难点问题,提出了一种基于神经网络集成的传感器故障诊断方法。该方法将传感器输出看做时间序列,通过加噪声生成抖动数据,建立多组神经网络,以获得神经网络集成预测器输出。通过将预测器输出与传感器实际输出相比较获取残差序列,获得基于残差序列的传感器偏置故障和漂移故障的辨识策略,实现传感器故障在线诊断。应用结果表明该方法可以提高神经网络的运算精度,从而快速准确地检测和分离传感器故障,辨识传感器故障类型以及故障发生的时间。

Research on Sensor Fault Diagnosis Based on Neural Network Ensemble Method

Aiming to solve the challenging problem of diagnosis for sensorbias and driftfaults,a novel approach of sensor fault diagnosis based on neural networkensemble method isproposed.The method is treating the sensor output as a time seriesand creating multiple sets of neural networks by adding noise to sensor output to generate jittered data to obtain the neural network ensemble forecasting output.By comparingthe forecasting outputs of the neural network ensemble and the actual valuesof sensors,the identification strategy based on the sequenceof remain residuals for sensorsbiasfault and drift fault is acquired and on-line sensors fault diagnosis are carried out.The application results indicate that the approach can improve the accuracy of neural networkand accurately detect,isolateand identify the fault.