Research On Aggregation Method Of Multi-source Heterogeneous Security Monitoring Data In Power Distribution Internet Of Things

With the accelerated development of the ubiquitous sensing technology of the power distribution Internet of Things,the massive security monitoring data presents the characteristics of multi-source heterogeneity.It helps to perceive and process the security situation of the power distribution Internet of Things.However,how to effectively aggregate and analyze massive multi-source heterogeneous security monitoring data and extract useful information from it is still a difficult problem.This thesis focuses on the aggregation model and algorithm of multi-source heterogeneous security monitoring data in the power distribution Internet of Things.Based on the adversarial domain adaptation network,an adversarial multi-source heterogeneous data aggregation model is designed to realize effective classification and aggregation of multi-source heterogeneous security monitoring data,and considering the heterogeneity of data feature space and label space at the same time,a dynamic heterogeneous data transformation strategy is further proposed,which improves the generalization and robustness of the data aggregation model.Finally,for the multi-source domain to address the difference in the degree of contribution,a fine-grained multi-source sub-network structure is designed to improve the accuracy of data aggregation.The main research contents include:(1)Multi-source heterogeneous data aggregation model based on adversarial domain adaptationThe data aggregation model based on traditional machine learning mainly uses a large amount of labeled data for training,and relies too much on high-quality labeling and strong computing power to build the model,which cannot be applied to the actual application process of the power distribution Internet of Things without labels to multi-source heterogeneous data.An adversarial multi-source heterogeneous data aggregation model based on adversarial domain adaptation network is designed.The heterogeneous features of the data are transformed in the feature extractor,and the multi-source data is classified and aggregated by the domain discriminator and the label classifier,which improves the adaptability of the data aggregation model based on machine learning to multi-source heterogeneous data.(2)Dynamic heterogeneous data transformation strategy based on heterogeneous domain adaptationThe security monitoring data of the power distribution Internet of Things not only has various modal structures,but is also constantly changing.At this time,not only the feature space of the data will be heterogeneous,but the label space will also be different.Considering the influence of heterogeneous feature space and heterogeneous label space on the data aggregation process,a dynamic heterogeneous data transformation strategy is designed to optimize the feature extractor and domain discriminator.Perform flexible conversion,and introduce category weight parameters and dynamic factors to solve the problem of category imbalance and different importance of data distribution,realize fine-grained distribution alignment of heterogeneous data,and improve the generalization and robustness of the model..(3)Fine-grained multi-source sub-network structure based on multi-source domain adaptationThe contribution of each source domain data to the target domain is often different.When a single domain discriminator and label classifier are applied to heterogeneous data from multiple source domains,it is difficult to achieve the ideal data aggregation effect.A fine-grained sub-network structure is designed to optimize the domain discriminator and label classifier.By constructing multiple fine-grained sub-network structures to align the data distribution between each source domain and target domain,and for the problem of different domain contributions,the domain weight parameter is introduced to enable the model to better utilize the highly correlated source domain,thereby avoiding negative transfer and improving the accuracy of the model.