基于轻量级卷积神经网络的GIS绝缘和机械故障诊断方法

Insulation and Mechanical Fault Diagnosis Method for GIS Based on Lightweight Convolution Neural Network

绝缘和机械故障是气体绝缘金属封闭开关设备(gas insulated metal-enclosed switchgear,GIS)中占比最大的故障类型,准确的故障诊断和状态评价对保证电力系统安全稳定运行具有重要意义。深度学习方法已成为故障诊断领域的主流,但传统卷积神经网络需要强大的计算资源,在计算能力一般的智能终端设备中难以应用。为此,文中提出了基于轻量级卷积神经网络的GIS绝缘和机械故障诊断方法。首先,采用空间可分离卷积代替传统卷积构造EffNet轻量级卷积神经网络,大幅度降低了模型的计算量;其次,采用迁移学习策略进行模型训练,在提升网络识别准确率的同时加快了模型的收敛速度,解决了现场数据不足的问题;最后,通过t分布随机邻近嵌入对卷积神经网络特征进行可视化,进一步验证了模型的性能。研究结果表明,具有5个EffNet卷积块的轻量级卷积神经网络模型在绝缘和机械故障诊断中的准确率分别达到94.7%和98.7%,并显著降低了参数量、存储空间和计算开销等,适用于GIS智能组件和检测仪器,是电力物联网下嵌入式系统和移动终端的最佳选择。

Insulation and mechanical fault is the type of faults with the largest proportion in gas insulated metal-enclosed switchgear(GIS).Accurate fault diagnosis and state evaluation are of great significance to ensure safe and stable operation of power system.Deep learning method has become the mainstream in the field of fault diagnosis,but traditional convolution neural network requires powerful computing resources and is hard to be used in the intelligent terminal devices with average computing ability.Therefore,a kind of insulation and mechanical fault diagnosis method for GIS based on lightweight convolution neural network was proposed in this paper.Firstly,the spatially separable convolution was used to replace the traditional convolution to construct the EffNet lightweight convolution neural network,which effectively reduces the calculation of the model.Secondly,model training was implemented by using transfer learning strategies,which speeds up the accuracy of network recognition and accelerates the convergence rate of the model at the same time,and the problem of insufficient field data was solved.Finally,the convolution neural network features were visualized by t-distribution random neighboring embedding and the performance of model was verified further.The research results show that the accuracy of the lightweight network model with 5 EffNet convolution blocks is 94.7%and 98.7%respectively in the insulation and mechanical fault diagnosis,which reduces remarkably the reference quantity,storage and computing overhead.The method proposed in this paper is suitable for GIS intelligent components and detection instruments,which is the best choice for the embedded system and mobile terminal in power internet of Things.