摘要
首先利用粗糙集理论对原始数据进行约简，并按一定的原则选取多个的简。然后在每一个约简的基础之上构建一个前馈子网络，并将多个子网综合成统一的容错前馈神经网络，达到对冗余信息的综合利用。通过对相应权值的训练调节，使网络的输出更精确合理。即，当某些量测信号丢失或难以获得时，可以通过其它不包含该量测信号的的简所构成的网络来进行正确的诊断，从而在信息不完备、不精确的情况下，仍保持较好的诊断性能。最后，通过对某液体火箭发动机泄漏故障检测的仿真，表明该容错网络可以满足高可靠性诊断场所的需要。
We propose a highly reliable strategy for fault diagnosis based on fault--tolerance neural networks. First, we derive some reductions from crude data based on rough sets theory and choose more than one reduction according to some criteria. Then, we build a general fault--tolerance neural networks by synthesizing each subnet which is decided by a reduction so as to make full use of the redundant information. Furthermore, we adopt BP algorithm to adjust the weights of the networks to obtain the more accurate output value. That is to say, when some measured signals are missed or difficult to obtain, we can still detect the faults correctly by using the subnets corresponding to the reductions that do not include those signals. Finally, we apply the neural networks to detect fuel leakage in a rocket engine and the simulation results illustrate the effectiveness of the approach.
出处
《电机与控制学报》
EI
CSCD
2000年第2期117-121,共5页
Electric Machines and Control
基金
国家自然科学基金资助项目(69904004)。
