神经网络结构与算法对变压器故障诊断的影响

Influences of structures and the algorithms of the neural network for transformer on the performance of the fault diagnosis

BP神经网络算法本质上是基于梯度下降的一种迭代学习算法,存在学习收敛速度慢、收敛精度低、易陷入局部极小、学习率难以选取、隐层数及隐层神经元个数难以确定等缺陷。为了选择出更适宜变压器DGA故障诊断的神经网络结构及算法,本文采用了常用的几种智能算法对变压器故障样本进行了诊断性能对比实验。结果得出Levenberg-Marquardt神经网络算法是收敛速度较快的算法,有动量和自适应的梯度下降法是收敛稳定性较佳的算法;网络最优结构设计过程,为用于变压器DGA故障诊断的神经网络的结构和算法提供了系统化的试验方法。

The algorithm of BP neural network is essentially an iterative learning based on gradient descent,there is slow learning convergence, low convergence accuracy, easy to fall into local minimum, the learning rate is difficult to select, hidden layers and hidden layer neuron numbers are difficult to determine defects. In order to select a more suitable structure and algorithm of neural network based on transformer DGA fault diagnosis,this paper uses several intelligent algorithm of transformer fault samples diagnostic performance comparison experiment. Results indicated that algorithm of LevenbergMarquardt neural network is faster algorithm,gradient momentum and adaptive convergence is stability better algorithm; network optimal structural design process,based fault diagnosis of transformer DGA method the neural network structure and algorithm provides a systematic test methods. The design process of the optimal structure provides systematic test method based on the structure of the neural networks and algorithm of transformer DGA fault diagnosis.