利用神经网络规则提取方法获取转静碰摩故障诊断知识

Rotor-stator Rubbing Fault Diagnosis Knowledge Acquisition Using Rule Extraction from Neural Networks

获取易于理解的航空发动机转静碰摩故障诊断知识规则对于深刻认识碰摩故障机理、有效诊断碰摩故障具有重要意义。鉴于此,研究了一种新的基于功能性观点的神经网络(NN)规则提取方法,介绍了连续属性离散、训练样本产生、神经网络训练、示例样本产生及规则提取等关键算法,并用Iris数据对方法进行了验证。最后,通过航空发动机转子实验器获取碰摩故障样本,利用神经网络规则提取方法从故障样本中提取了碰摩故障诊断知识规则,并对其进行了验证分析,验证结果充分表明了该方法的正确有效性。

It is very important to acquire the easily understood diagnosis knowledge rules of rubbing fault, in order to further understand the rubbing fault mechanism and effectively diagnose the rubbing fault. In this article, a rule extraction method based on the functional point of view is studied, and the key algorithms are introduced, such as the discretization of continuous attributes, the generation of train samples of neural network （NN）, the training of NN, the generation of instance samples from the trained NN, and the rule extraction. The Iris dataset is used to verify the rule extraction method. Finally, rotor-stator rubbing fault samples are obtained by an aero-engine rotor experimental rig, the rule extraction method is used to extract the rubbing fault diagnosis knowledge rules from fault samples, the obtained rules are verified and analyzed, and the results fully show the correction and rationality of the new method.