Research On Transformer Fault Diagnosis Based On Multi-source Heterogeneous Data Fusion

Transformer is a vital infrastructure in the field of electric power in our country,related to power transmission,voltage and current conversion,energy loss and many other energy problems.The health state of transformer is related to the normal operation of local large power network.The traditional industry of transformer fault judgment,is the use of manual judgment,the need to shut down the transformer for maintenance,the maintenance operation is affected by the temperature environment,electromagnetic environment,humidity,resulting in the manual maintenance process cumbersome,consuming time,high maintenance cost,fuzzy discrimination standards.In this paper,the fault condition of the fan inside the transformer is determined by collecting the sound data of the transformer,and the abnormal phenomenon of DC magnetic bias and discharge is determined.The spectrogram corresponding to the sound signal is generated and the convolution neural network is combined to accurately identify and judge the fault condition and typical faults of key components.Finally,for the discharge fault,the dissolved gas in the oil data is collected.Based on the fuzzy theory,the improved three-ratio method is used to judge the fault types of fine discharge such as low energy discharge,arc discharge and partial discharge,and to judge the minor overheating fault types such as low temperature overheating,medium temperature overheating and high temperature overheating.Compared with the traditional data using a single source,such as sound signal or dissolved gas data in oil,In this paper,multi-source heterogeneous data is used for information fusion to comprehensively diagnose transformer health status.It has the advantages of high accuracy,good model generalization ability,high robustness,high credibility and strong discrimination basis.Compared with the manual maintenance method,the convolutional neural network and fuzzy theory intelligent transformer fault judgment technology adopted in this paper has the advantages of fast maintenance speed,high identification accuracy and low judgment cost.The main work of this paper is as follows:(1)Professional instruments are used to collect the transformer sound data in the substation without noise,and the standard sample library is made by mixing common environmental sounds to simulate the real working environment of the transformer.After MFCC spectrogram generation technology,it is observed that the energy density of spectrograms of different fault types is different at different frequencies.The difference between different faults is the theoretical basis for the recognition of different fault types by convolutional neural networks.In summary,the feature extraction algorithm based on fault spectrogram is researched.(2)Based on spectrogram feature data,extract sound feature maps of different fault spectrograms,build convolutional neural network model in Pytorch tool environment,use convolutional neural network model to learn the subtle differences between spectrograms corresponding to different faults,and further improve the performance of CNN network model by adjusting model parameters and model optimization.So as to achieve the identification accuracy with industrial application value.(3)Based on the IEC improved three ratio algorithm of dissolved gas in oil,combined with the study of fuzzy theory to overcome the shortcomings of the traditional three ratio method,the improved algorithm is proposed.The example is verified that the improved algorithm has higher judgment accuracy than the original algorithm,and combined with CNN,it can fully deduce the primary and secondary fault types to be refined.