基于域泛化的非均衡电力设备分/合闸X射线图像识别

Imbalanced power equipment opening/closing X-ray image recognition based on domain generalization

针对电力开关设备分/合闸X射线图像呈现设备种类多、数据量少、类别不平衡问题,提出一种基于域泛化的非均衡电力设备分/合闸X射线图像识别方法。首先以AlexNet和改进的ResNet作为骨干网络设计识别算法;然后利用最大熵对抗数据增强(MEADA)训练算法对训练集进行样本扩充,以模拟未知类型分/合闸图像的域分布;最后通过将Focal Loss损失函数引入到识别算法中解决分/合闸数据类别不平衡问题。实验结果表明,使用所提方法AlexNet和简化ResNet(ResNet-F)模型的平均准确率相比其基线方法分别提升5.31个和6.52个百分点,且ResNet-F的识别精度比AlexNet高出3.54个百分点。类激活图、受试者工作特征(ROC)曲线和t-随机嵌入(t-SNE)等可视化分析结果进一步验证了所提方法的有效性,为多域非均衡电力设备分/合闸X射线图像识别提供了新思路。

Aiming at the problems of many types of equipment,small amount of data,and imbalanced class distribution in power switchgear opening/closing X-ray images,an X-ray image recognition method for opening/closing of imbalanced power equipment based on domain generalization was proposed. Firstly,the recognition algorithm was designed with AlexNet and improved ResNet as the backbone network. Then,the training algorithm based on Maximum-Entropy Adversarial Data Augmentation(MEADA)was used to expand the training set to simulate the domain distribution of unknown type opening/closing images. Finally,the Focal Loss function was introduced into the recognition algorithm to solve the problem of opening/closing data class imbalance. Experimental results show that compared with the baseline method,the average accuracy of AlexNet and ResNet-F models using the proposed method is improved by 5. 31 and 6. 52 percentage points respectively,and the recognition accuracy of ResNet-F is 3. 54 percentage points higher than that of AlexNet. The effectiveness of the proposed method is further verified by the visualization analysis results of Grad-CAM(Class Activation Map),ROC(Receiver Operating Characteristic)curve and t-SNE(t-distributed Stochastic Neighbor Embedding),which provides a new idea for multi-domain imbalanced power equipment opening/closing X-ray image recognition.