采用改进分布式模型的传感器故障隔离方法

Sensor fault isolation method using improved decentralized model

针对基于分布式模型的传感器故障隔离方法中传统模型结构的不足,从模型结构和建模精度两个方面提出一种改进方法.首先针对传统模型结构直接将传感器输出作为模型输入,易导致故障虚警的问题,给出一种时间窗交互预测的改进结构;然后采用神经网络作为建模工具,设计了结合粒子群算法和梯度法的网络训练方法,以克服梯度法局部最优的缺陷.以B747飞机模型为诊断对象,通过分别设置攻角传感器和俯仰角速度传感器发生漂移或恒偏差故障,对比验证了所提方法的有效性.

Aiming at the structural defect of the sensor fault isolation method based on decentralized model,an improved method in both model structure and precision was proposed.Firstly,a novel model structure based on interactive prediction in time window length was designed to replace the traditional structure,which uses sensor outputs directly as model inputs and may trigger false alarms easily.Then with some neural network taken as the modeling tool,a hybrid training strategy combining particle swarm optimization（PSO）and gradient method was presented to deal with the local optima shortcoming of the gradient method.Based on the simulation model of the B747 airplane,the angle of attack sensor or pitch rate sensor were set with drift and fixed bias faults,respectively.The obtained contrast results validate effectiveness of the method.