| 摘要: |
| 为了实现模块化多电平变换器(MMC)子模块开关管的故障诊断，提出了一种基于挤压-激励模块的深度可分离卷积神经网络(SE-DSCNN)。该网络直接利用原始电容电压数据，不需要任何的特征提取算法，能够自动提取隐藏在原始数据中的深层特征，结合挤压-激励模块以突出通道域中具有代表性的特征，利用深度可分离卷积(DSC)来减少网络的计算量。利用滑动时间窗口将数据分段并归一化后输入提前训练好的最优模型中，模型输出预测标签。通过与其他人工特征提取方法及深度学习方法进行对比，结果表明模型参数量比具有相同卷积层数的标准卷积神经网络(CNN)减少了70.92% 左右。所提方法在已有样本片段上的分类准确率及不同故障时期的诊断正确率均达99 % 及以上，诊断单个样本片段所需的时间约为0.34 ms，不但能区分故障早期的耦合性特征，还能实现准确、可靠、高效、快速的故障诊断。 |
| 关键词: MMC 开关管故障 挤压-激励模块 深度可分离卷积神经网络 故障诊断 |
| DOI：10.16081/j.epae.202202029 |
| 分类号:TM46 |
| 基金项目:国家自然科学基金资助项目(51977153,51977161,51577046)；国家自然科学基金重点项目(51637004)；国家重点研发计划“重大科学仪器设备开发”资助项目(2016YFF0102200)；装备预先研究重点项目(41402040301)；武汉市科技计划项目(20201G01) |
|
| Fault diagnosis method for switches in MMC based on SE-DSCNN |
|
ZENG Zhaorong, HE Yigang
|
|
School of Electrical Engineering and Automation, Wuhan University, Wuhan 430072, China
|
| Abstract: |
| In order to realize the fault diagnosis for switches in MMC(Modular Multilevel Converter),a SE-DSCNN(Depthwise Separable Convolutional Neural Network based on Squeeze-and-Excitation) module is proposed. The network directly uses the original capacitance voltage data without any feature extraction algorithm, and can automatically extract the deep features hidden in the original data. It highlights representative features through the combination of SE module, then reduces the calculation of network through DSC(Depthwise Separable Convolution). A sliding time window is used to segment and normalize the data that is then fed into the optimal model trained in advance, and then the model outputs the prediction label. Compared with other manual feature extraction methods and deep learning methods, the results show that the amount of model parameters is reduced by about 70.92% compared to the standard CNN(Convolutional Neural Network) that has the same number of convolutional layers. The classification accuracy on the existing sample fragments and the diagnosis accuracy between different fault periods for the proposed method are both 99 % and above. The time required to diagnose a single sample fragment is about 0.34 ms. It can not only distinguish the coupling characteristics of early failures, but also realize accurate, reliable, efficient and fast fault diagnosis. |
| Key words: MMC switch failure squeeze-and-excitation module depthwise separable convolutional neural network fault diagnosis |
