基于模糊输出BP神经网络的电机主绝缘老化状态评估方法

评估大电机主绝缘的老化状态是电气工程领域中的重要课题，其目标是根据非破坏性参量评估绝缘的老化状态，然而传统的评估方法一直没有得到准确可靠的评估结果。为了克服传统的绝缘状态阈值评估方法的缺点，提出应用神经网络和模糊数学结合评估绝缘状态的新方法。由于绝缘的老化状态是一种模糊现象，所以建立了具有模糊输出的3层BP神经网络对绝缘状态进行评估，并确定了绝缘状态论域上的四个模糊子集及其隶属函数，同时规定了网络的输入和输出参量。选择Levenberg-Marquardt快速学习算法对该文建立的模糊输出神经网络进行了训练，并使用另外10组多因子加速老化数据和5组真机线棒数据对网络的评估能力进行了检验，检验结果表明该方法可以准确有效地评估定子绝缘的老化状态。

EVALUATING THE AGING CONDITION OF MAIN INSULATION IN LARGE GENERATOR BASED BP ARTIFICAL NEUTRAL NETWORK WITH FUZZY OUTPUT

The condition assessment of large generator main insulation is the important research subject in electrical engineering field, and the key destination is to assess the aging condition of insulation based on the nondestructive parameters. However, the exact assessment result hasn't been acquired until now according to traditional assessment methods. A new assessment method based on the combination of fuzzy math theory and artificial neutral network (ANN) is proposed in order to overcome the disadvantages of traditional insulation condition assessment based on the threshold model. As the aging condition of insulation is a kind of fuzzy phenomenon, the 3 layers of BP ANN are established with 4 fuzzy outputs, which are the degrees of membership to four fuzzy subsets of insulation condition respectively, and 28 inputs corresponding to 28 nondestructive parameters of insulation respectively. The ANN with fuzzy outputs is trained by the Levenberg-Marquardt fast training algorithm with the goal error of 0.0001. Finally, the ability of condition evaluation of the network is verified by ten stator bars subjected to multi-stress accelerated aging and five 18kV/300MW stator bars. The verification results show that the method could assess the aging condition of stator bar insulation effectively and accurately.