云边协同下基于深度迁移网络的配电台区异常工况诊断方法

Diagnosis method of abnormal working conditions in distribution station area based on deep migration network under cloud-edge collaboration

为了实现配电台区异常工况精细化诊断,提出云边协同下基于深度迁移网络的配电台区异常工况诊断方法。首先,在云中心对多个相似配电台区的异常工况样本进行汇集,利用精细化的运行工况样本集训练构建源域异常工况诊断的卷积神经网络模型。其次,将源域诊断模型迁移至目标域的单一配电台区边缘节点处,利用迁移机制进行目标域上的差异性训练,引入多核最大均值差异来计算源域与目标域的分布差异,构建目标域优化损失函数,使目标域与源域自适应匹配,从而有效建立目标域异常工况诊断模型。通过实验验证所提方法具有良好的异常工况精细化诊断能力,诊断性能明显优于其他常规方法。同时,该方法能减缓云中心集中训练诊断模型的计算资源需求压力,有效利用边缘节点的计算能力和响应能力。

In order to realize the fine diagnosis of abnormal working conditions in the distribution station area, a method for diagnosing abnormal working conditions in distribution station area based on deep migration network under cloud-edge collaboration was proposed. Firstly, abnormal working condition samples from multiple similar distribution station areas was collected in the cloud center, and the refined operating condition sample set was used to train and construct a convolutional neural network model for abnormal working condition diagnosis in the source domain. Secondly, the source domain diagnosis model was migrated to the edge node of a single distribution station area, and the differential training on the target domain was performed by using the migration mechanism. Multi-kernel maximum mean discrepancy was used to calculate the distribution difference between the source domain and the target domain, and the optimization loss function of the target domain was constructed to make the target domain and the source domain adaptively match, so as to effectively establish the abnormal working condition diagnosis model in the target domain. Through the experiments, it was verified that the proposed method had good ability of fine diagnosis of abnormal working conditions, and the diagnostic performance was significantly better than other conventional methods. At the same time, this method can alleviate the pressure of computing resources in the cloud center to train diagnostic models, and effectively utilize the computing power and response ability of edge nodes.