Abstract: In recent years,intelligent fault diagnosis methods based on deep learning technology have be.come a research hotspot in the field of electrical converters.Convolutional Neural Network (CNN) has the potential to identify early-term micro faults because of its powerful feature extraction capabilities.The current Convolutional Neural Network (CNN) algorithms are not suitable for the rapid diagnosis of electrical faults due to its complicated model structure,excessive training parameters and long diagnosis waiting time.For this reason,a new deep learning algorithm based on improved.cnN-GAP-SVM was proposed for the rapid diagnosis of DC-DC converters early faults.The model input layer automatically transformed the raw 1D time-series monitoring data into 2D feature map training samples.To reduce model parameters,a Global Average Pooling Layer (GAP) was designed to improve the fully connected network structure of 2~3 layers in the traditional.cnN.To further improve the diagnosis accuracy,a nonlinear Support Vector Machine (SVM) was used to replace the traditional Softmax function as the final classifier.Experiments showed that the proposed method could improve the diagnostic accuracy to 100% and the diagnosis speed by 23%.compared with the traditional diagnosis methods,the proposed method had faster diagnosis speed and higher diagnosis accuracy.

Key words: intelligent fault diagnosis, convolutional neural network, support vector machine, DC-DC converter, open-circuit fault

摘要： 近年来,基于深度学习技术的智能故障诊断方法在电力变换器领域得到了广泛研究。卷积神经网络(CNN)因其强大的特征提取能力而具备辨识早期微小故障的潜力。然而,现行的CNN算法因其模型结构过于复杂、训练参数量较多、诊断时间较长而不适用于电气故障的快速诊断。为此,提出了一种基于改进CNN-GAP-SVM的深度学习新算法,用于DC-DC变换器早期故障的快速诊断。首先,将原始的时间序列监测数据转变为二维特征图故障样本;其次,该方法设计了一个全局均值池化(GAP)层,用于代替传统CNN中2~3层的全连接层部分,以减少模型参数量;然后,采用非线性支持向量机(SVM)代替传统Softmax函数作为最终分类器,进一步提升诊断精度。实验表明;所提方法不仅将诊断准确率提升至100%,还提升了23%的诊断速度。通过与传统智能诊断方法相比较,证明了所提方法具有更快的诊断速度和更高的诊断准确率。

关键词: 智能故障诊断, 卷积神经网络, 支持向量机, DC-DC变换器, 开路故障