基于改进LSTM的电抗器故障预警方法

Reactor fault early warning method based on improved LSTM

针对传统的电抗器故障预警方法存在过拟合以及缺乏通用性的问题,提出了基于长短期记忆神经网络(LSTM)和孪生神经网络的电抗器故障预警方法(以下简称HDDse)。该方法结合长短期记忆神经网络和孪生神经网络的优势,LSTM结构用于学习电抗器健康状态下的动态变化行为,孪生网络结构可以降低电抗器信息映射到高维空间的学习效率,该方法已成功地应用于电抗器的故障诊断中。实验结果表明,与已有电抗器故障预警方法相比,HDDse可以极大提升电抗器故障预警准确性。

To solve the shortcomings of over-fitting and lack of generality of traditional reactor fault early warning methods,a reactor fault early warning method based on short-term and short-term memory neural network and twin neural network(hereinafter referred to as HDDse)is studied.This method combines the advantages of long and short term memory neural network and twin neural network.LSTM structure is used to learn the dynamic change behavior of reactor under healthy state.Besides,twin network structure can reduce the learning efficiency of reactor information mapping to high-dimensional space.This method has been successfully applied in reactor fault diagnosis.The experiment results show that HDDse can greatly improve the accuracy of reactor fault early warning compared with the existing reactor fault early warning methods.