一种能给出充油电气设备油色谱故障诊断可靠性的神经网络方法

为了提高基于人工神经网络方法的充油电气设备油色谱故障诊断的准确性及诊断结果的可靠性,基于神经网络理论分析指出了采用不同训练算法、隐层神经元数量、初始权值和阈值训练得到多个网络输出的均值作为诊断结果能提高故障诊断的准确性,根据多个网络输出的标准差可以获得诊断结果的可靠性.根据搜集得到的大量油色谱样本,分别采用振荡传播(resilient propagation,RPROP)算法、共轭梯度法、拟牛顿法和Levenberg-Marquardt算法训练共计得到40个结构相似的神经网络,将训练得到神经网络应用于基于油色谱的充油电气设备故障诊断,同时比较了不同算法的训练时间和诊断结果的准确性.结果 表明多个网络输出的平均可提高故障诊断的准确性,根据多个网络输出的标准差可获得诊断结果的可靠性,而且表明神经网络结构相似时,4种算法训练得到的神经网络具有相近的故障诊断准确性,但从训练时间上看,RPROP算法、拟牛顿法和Levenberg-Marquardt算法非常接近,而共轭梯度法的训练时间为其他3种算法的6倍左右.同时考虑到Levenberg-Marquardt算法计算速度最快,可在充油电气设备油色谱故障诊断中用于训练神经网络.

An ANN method for oil-filled electrical equipment fault diagnosis with reliability based on DGA

To improve the accuracy and reliability of dissolved gas-in-oil analysis (DGA) fault diagnosis of oil-filled electrical equipment based on artificial neural network (ANN) method, based on the analysis of ANN theory, it is pointed out that averaging the output of multiple ANNs trained with different algorithms, number of hidden neurons, initial weights and thresholds can improve the accuracy of fault diagnosis, and at the same time, the reliability of the diagnosis results can be obtained according to the standard deviation of the output of multiple ANNs. According to a large number of collected DGA samples, 40 ANNs with similar structure were trained by the RPROP (Resilient Propagation) algorithm, the conjugate gradient method, the quasi-Newton method and the Levenberg-Marquardt algorithm respectively. The trained ANNs were applied to the fault diagnosis of oil-filled electrical equipment based on DGA. At the same time, the training time and the accuracy of fault diagnosis by the ANNs trained by different algorithms were compared. The results reveal that averaging the output of multiple ANNs can improve the accuracy of fault diagnosis, and at the same time, the reliability of the diagnosis results can be obtained according to the standard deviation of the output of multiple ANNs. The ANNs trained by different algorithms have similar fault diagnosis accuracy. However, the training times of the RPROP algorithm, quasi-Newton algorithm and Levenberg-Marquardt algorithm are quite similar, the training time of the conjugate gradient algorithm is about 6 times of them. Considering that the Levenberg-Marquardt algorithm requires the least computational effort, it can be used to train ANNs for DGA fault diagnosis of oil-filled electrical equipment.