Condition Identification Of Power Transformer Core And Winding Based On Vibration Data Enhancement

In the context of the gradual promotion of strong smart grid construction,the requirements for intelligent operation and maintenance of electrical equipment,data mining and analysis are increasingly urgent.Deep learning methods are driven by large amounts of data and can effectively solve problems such as classification and prediction,providing new ideas to realise intelligent operation and maintenance of electrical equipment,which is of great value to improve the safety of power systems.Power transformers are the core equipment for energy conversion and transmission in power grids,and their vibration signals contain rich information,which is of great advantage for identifying mechanical-type faults such as loose cores and deformed windings.Therefore,the main work of this paper is to combine the vibration mechanism of power transformers,vibration data and deep learning methods,to study the preprocessing methods of vibration data,and to seek new ways in vibration data enhancement to achieve the condition recognition of transformer core windings.The work completed in the thesis is as follows.Based on the vibration mechanism of power transformer cores and windings,the typical characteristics of transformer vibration signals are analyzed,and the influence of transmission path on signal attenuation during signal propagation is explored in conjunction with vibration acoustics.The results show that the vibration signals on transformer cases can reflect the state of cores and windings,and the vibration signal acquisition of transformers in operation is carried out with this as a guide to verify the correctness of the theoretical analysis.The collected power transformer vibration signals were pre-processed with improved wavelet thresholds,resulting in a significant improvement in the de-noising effect.In order to realize the separation of transformer core and winding vibration signals,a transformer vibration signal separation method is proposed.FastICA is combined with VMD to achieve signal separation,and the effectiveness of the separation method is verified by simulated vibration signals as well as measured vibration signals.The value density of transformer vibration data is very low and cannot be efficiently utilized.In order to solve the problem of small sample data in transformer vibration data acquisition,a data enhancement method based on generative adversarial networks is proposed.A self-attentive mechanism is added to the gradient penalty-based Wasserstein generative adversarial network,and its hyperparameter selection method is improved to further enhance the applicability of the network model to vibration signals.A one-dimensional self-attentive Wasserstein generative adversarial network with gradient penalty is constructed for transformer vibration signals,and three types of fault signals,namely core loosening,winding deformation and winding loosening,are generated.The results show that the generated samples are of high quality in the time domain,frequency domain and data distribution,indicating that the model has a strong sample generation capability and is effective in improving the network hyperparameters.Based on the enhanced vibration data,a onedimensional convolutional neural network was constructed for state identification of transformer cores and windings.The network model was trained and tested using samples of different data sizes.The results show that the data enhancement method can well solve the problem of unbalanced vibration data classes,improve the availability of small sample data and increase the accuracy of the classification model.