Research On Intelligent Fault Diagnosis Method Of Traction Transformer Based On Noise Analysis

In recent years,with the rapid development of rail transit system,the requirements for the safe and stable operation of its traction power supply system are becoming more and more strict.Traction transformer is the key equipment of rail transit traction power supply system because it is responsible for the important functions of voltage conversion and power distribution.Therefore,the effective diagnosis and monitoring of transformer working conditions is of great significance to the safe and reliable operation of the whole rail transit system.Due to the characteristics of many types,complex internal structure and various operating conditions of transformers,it is difficult to detect their abnormalities.The transformer fault intelligent diagnosis based on noise analysis is a simple and effective non-invasive monitoring method,which will not affect the normal operation of transformers and can well reflect various operating conditions of transformers.Therefore,according to the different characteristics of noise signal in time and frequency,this paper discusses and studies various methods of four modules:preprocessing,feature analysis,feature extraction and classification and recognition in the intelligent fault diagnosis system of traction transformer based on noise analysis,and improves the classification and recognition algorithm based on deep learning according to the characteristics of rare fault noise,complex interference noise and large scale dimension of feature data of traction transformer noise signal,To obtain more effective and stable fault diagnosis performance.The specific work of this paper can be summarized as follows:1.In the preprocessing module of the fault diagnosis system,this paper uses wavelet threshold denoising as the filtering and denoising method of the noise signal of the traction transformer,adjusts and compares different parameters such as wavelet base,decomposition layers and threshold selection rules of the wavelet threshold denoising,and obtains the wavelet threshold denoising parameters suitable for the noise signal of the traction transformer through the parameter adjustment optimization experiment,so as to filter the interference noise to the greatest extent,Achieve better feature extraction effect.2.In the feature analysis module,various time-frequency analysis methods are analyzed and compared.The time-frequency characteristics of transformer noise under various working conditions are jointly analyzed by using the sound spectrum obtained by short-time Fourier transform combined with time-domain diagram and spectrum diagram.The time-frequency joint domain analysis can clearly describe the changing relationship of noise signal frequency at any time,which can better characterize the characteristics of transformer noise signal under different working conditions.3.In the feature extraction module,various time-domain features,frequency features and cepstrum features are analyzed and compared,and various features are visualized through experiments.Then,according to the characteristics of traction transformer noise signal,a scheme of using MFCC,LPCC and their combination as feature vector input to the classification and recognition module is proposed to achieve better classification and recognition effect.4.In the classification and recognition module,various machine learning algorithms are analyzed and compared.According to the characteristics of traction transformer noise signal,the improved CNN-LSTM model based on deep learning algorithm is studied.Through parameter adjustment experiment,input feature comparison experiment and classification and recognition algorithm comparison experiment,it is verified that on the traction transformer noise sample data set collected and used in this paper,the effectiveness of the model in intelligent diagnosis of noise fault of traction transformer.