基于多级RBF神经网络集成的传感器故障诊断研究  

Research on Sensor Fault Diagnosis Based on Multistage RBF Neural Network Ensemble

作　　者：李明[1] 税爱社[1] 宋政辉[1] 李林[2] 李智宇[3] 

机构地区：[1]后勤工程学院后勤信息工程系,重庆401311 [2]78479部队,成都610683 [3]73891部队,福州364000

出　　处：《后勤工程学院学报》2010年第4期76-81,共6页Journal of Logistical Engineering University

基　　金：重庆市自然科学基金资助项目(CSTC,2007BB2101)

摘　　要：针对诊断传感器偏置故障及漂移故障的难点问题，提出了一种基于多级RBF神经网络集成的传感器故障诊断方法。该方法充分利用控制系统闭环回路测控信息，建立多级神经网络集成观测器模型。将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，实现控制系统传感器故障在线诊断。将三容水箱液位控制系统作为仿真对象，仿真结果表明该方法不仅可以提高单一神经网络的运算精度，而且采用RBF神经网络集成方式还要优于其他集成方式，可以快速准确地检测和分离传感器故障，辨识传感器故障类型、故障大小以及故障发生的时间。Aiming at the challenging problem of diagnosis for sensor bias and drift faults,a novel approach of sensor fault diag- nosis based on multistage RBF neural network ensemble is proposed, by which the closed-loop monitoring information in control system is adopted to establish multistage RBF neural network ensemble observer. By comparing the outputs of multistage RBF neural network ensemble observer and the actual values of sensors, the identification strategy based on the sequence of remain residuals for sensors bias fault and drift fault is acquired and on-line sensors fault diagnosis in control system are carried out. The three tank system simulation results indicate that the approach can improve not only the accuracy of neural network, but also the neural network ensemble method and accurately detect, isolate and identify the fault.

关 键 词：神经网络集成 观测器 传感器 故障诊断 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程] TP212[自动化与计算机技术—控制科学与工程]