摘要
为解决传统神经网络进行传感器故障诊断时存在的过拟合、泛化能力有限等问题,提出一种基于深度置信网络观测器的航空传感器故障诊断方法。利用深度置信网络替代浅层神经网络,在优化网络结构的基础上,给出深度置信网络隐层节点数选取的递推公式,构建深度置信网络状态观测器。离线训练时,利用飞行数据训练深度置信网络观测器。在线诊断时,通过比较观测器输出值与实际输出值判断故障类型,并给出3种故障隔离与信号重构方法。仿真结果表明,与BP神经网络观测器相比,该方法能够快速准确地进行故障诊断与隔离,并且完成信号重构。
To solve the problem of over-fitting and limited generalization ability during sensor fault diagnosis by traditional neural network,a fault diagnosis method for aerial sensor based on deep belief network observer is proposed.Shallowlayer neural network is replaced by deep belief network. On the basis of optimizing network structure,the recurrence formula of selecting hidden layer nodes is proposed to build deep belief network state observer. Flight data is used to train deep belief network observer during offline training. Output of the observer is compared with actual output to judge the fault types and three methods of fault isolation and signal reconstruction are proposed during online diagnostics.Simulation results showthat compared with BP neural network observer,the proposed method can diagnose and isolate faults and reconstruct signals with rapidity and high accuracy.
出处
《计算机工程》
CAS
CSCD
北大核心
2017年第7期281-287,共7页
Computer Engineering
关键词
航空传感器
故障诊断
深度学习
深度置信网络
故障隔离
信号重构
aerial sensor
fault diagnosis
deep learning
deep belief network
fault isolation
signal reconstruction
