| The power transformer is the core equipment in the power delivery system,and its operating condition will affect the security and stability of power grid directly.In order to find potential faults timely and effectively,the predecessors integrated rich structured data to assess the components and the overall operating condition.However,during the operation and maintenance of power transformer,the inspection personnel recorded unstructured text information such as fault locations,fault performance,and maintenance measures,which was stored in work tickets and operation tickets.This information not only reflected the historical trend of operating condition,but also contained various potential fault information,which is an important supplement to other structured data.However,due to the complexity of text structures and the ambiguity of text semantics,text data are still untapped.In order to carry out condition maintenance for monitoring and assessing the operating risk comprehensively,it is necessary to establish a text-mining model,and fuse structured and unstructured data.The article researched the information mining models corresponding to different structural data.For information mining of operation and maintenance texts,text preprocessing methods such as text segmentation,removal of stop-words,and generation of distributed word embeddings were proposed to obtain high-quality input sets.Besides,a text-mining model based on deep semantic learning was built.The model can extract deep semantics based on its end-to-end architecture.Meanwhile,the text characteristics were summarized and targeted model optimization measures were proposed.For information mining of structured data,the characteristic indexes that can reflect the transformer operating condition were extracted.Considering the correlation between equipment condition and time,a dynamic data analysis model based on time series characteristics was proposed.Besides,conditional entropy was introduced and time-dependent analysis of related data was given.For overcoming the insufficient number of fault data samples,an effective self-learning strategy for fault data was proposed based on transfer learning.In addition,a model about heterogeneous data fusion was also proposed.Through the softmax classification layer,the probability distribution of equipment failure levels was output.Combined with the State Grid condition maintenance guideline,the health indexes of unstructured texts and structured data were obtained respectively.Besides,a time sequence diagram of transformer condition change portrayed by the health index was proposed,and the unit health cycle was introduced to describe the periodic dynamic process of the fault event.Moreover,a proportional failure rate model was established.It can estimate the failure rate using partial likelihood theory,and obtain the fault occurrence rate and component degradation rate at different times.Thereby the potential development trend can be predicted.The case based on actual recorded data of the power grid shows that the index characterizing the model performance increase by 2.64%-6.13% after adopting targeted optimization strategies.Besides,the feature vectors projected into the two-dimensional space can distinguish categories clearly,and the model has an excellent feature extraction capability.In terms to the structured data analysis,the classification accuracy rate is 91.67%,while in terms to the analysis of structured data and unstructured data,the accuracy rate is 96.67%.Compared with the former,the latter has an accuracy rate increased by 5%,which verifies the effectiveness of the proposed identification system of failure risk.Meanwhile,the life-time assessment time chart can not only describe the health index of each moment effectively,but also reflect important fault information accurately,including the fault occurance time,the degradation degree,the fault handle time,and the latent fault.It is an effective auxiliary tool for identifying equipment failure risk. |
PDF Full Text Request