Research On Typical Defects Recognition Of Dry-type Transformer Based On Voiceprint

As an important power equipment in the distribution network,dry-type transformers may cause faults once they have defects,which will cause huge losses to the system.The sound signal(voiceprint)of dry-type transformers under normal operating conditions and typical defect states has obvious characteristic differences.Therefore,the characteristic analysis and pattern recognition of the dry-type transformers voiceprint has become important in transformers online monitoring field.This work has great practical significance for transformer state warning and fault diagnosis.In this paper,a dry-type transformer typical defect simulation test platform and voiceprint acquisition device are built.Under the test conditions of low background noise,the transformer core loosening,winding loosening,winding deformation,core multi-point grounding,partial discharge failure,overexcitation,and overloading simulation tests were carried out.The collection of the voiceprint signal is completed by the voiceprint acquisition device.Through analyzing and studying the voiceprint signal under normal operating state and typical defect state,the frequency domain distribution of voiceprint signal under different working conditions is obtained.Then,the main frequency,odd and even harmonic ratios and high and low frequency ratios are taken as characteristic values.The calculation results show that the main frequency of the voiceprint signal is shifted from 100 Hz to high frequency when the iron core clip looseness and the winding clip looseness occur,and the high and low frequency ratio is reduced;the partial discharge voiceprint has obvious high-frequency ultrasonic components;the harmonic component of the voiceprint signal under the over-excitation state increases sharply;while the iron core multi-point grounding,slight winding deformations and overload conditions have no obvious characteristics.The frequency domain distribution and characteristic values can only distinguish partial defects from normal conditions,but it is difficult to accurately identify the types of defects,which highlights the advantages of deep learning for the identification of typical defects of dry-type transformers.Finally,the Mel-CNN recognition method is proposed,the original data is filtered by Mel filter,and the Mel-spectrum is obtained as the recognition data set,which realizes the feature extraction and dimensionality reduction of the voiceprint.The structure adjustment and the hyperparameter optimization are performed,which can effectively identify typical defects of dry-type transformers.The research results show that the accuracy of typical defect recognition reaches 98.53%,and the recognition results verify the feasibility of the method,which provides a reference for data mining and defect recognition of dry-type transformers.