利用神经网络进行智能诊断所获取的诊断知识难于解释和理解,因此限制了其在智能诊断中的进一步应用,基于此,本文研究了一种新的基于功能性观点的神经网络规则提取方法,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。并与其他方法进行了比较分析。最后,将方法应用于转子故障诊断,通过多功能转子故障模拟实验台获取故障实验样本,利用本文神经网络规则提取方法从故障样本中提取了诊断知识规则,并对其进行了解释,结果表明了方法的正确有效性。
In neural network intelligent (NN) diagnosis, the knowledge which is acquired by NN, is very difficult to be explained and understood, therefore, the further application of NN intelligent diagnosis is limited. In this paper, the rules extraction method from neural network 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 the instance samples from the trained NN, and rule extraction. The new method is compared with the other methods, and its correction is verified. Finally, the fault experimentation samples are obtained by a multi-function rotor experimental rig, the rules extraction method is used to extract the diagnosis knowledge rules from fault samples, and the results show the correctness and rationality of the new method.