Research On Stator Slot Wedge Loosening Detection Of Hydro-Generator Based On MFCC And CC Neural Network

China has incorporated "double carbon" into the national development strategy,and the power system is the cornerstone of ensuring the high-speed operation of the electrification era.Hydropower is at the backbone of low-carbon clean energy,while large hydroelectric generators play a key role in hydropower systems.High-quality maintenance keeps the generator in good condition,ensuring that it can operate stably and efficiently.The maintenance of the stator wedge is the key to ensuring the normal operation of large hydroelectric generators.In this paper,the tightness detection of stator trough wedges in the offline state of large hydroelectric generators is mainly studied.The main research work in thesis is as below:(1)An automated electromagnetic tapping device and a generator maintenance robot based on the orbital mechanism are proposed.According to the force of the stator rod in the magnetic field in the slot wedge,the degree of loosening of the stator wedge is divided into three types: tightening,slightly tightness and loosening.Then,under laboratory conditions,three different states of the stator slot wedge model is simulated,and the surface of the stator slot wedge is struck by electromagnetic percussion device,and the percussion sound signal in different states is obtained,each group of 100 percussion sound signal samples,a total of 300 raw sample data.(2)Preprocessing of raw sample data.The situation in the generator maintenance site is complex,and there is more noise interference.When tested under laboratory conditions,it is possible to choose whether to reduce the sound signal of the acquired percussion sample according to the working environment.Multi-window spectral estimation uses multiple data windows to calculate the direct spectrum,and then average,weakening the variance between adjacent frames,and the spectral subtraction method of multi-window spectral estimation can be used to reduce the noise of the sample data.At the same time,based on the digital signal entropy ratio method,the sample data of the original 3s is intercepted into a valid data fragment of 0.5s.This not only reduces the interference of external sound sources on the sample data but also increases the rate of classification operations.(3)Extracts the feature matrix of the effective fragments of the percussion sound signal.Using speech signal features,the feature parameters of the preprocessed sample data are extracted,and the Linear Prediction Cepstral Coefficients(LPCC)and Mel Frequency Cepstral Cofficients(MFCC)feature matrices of 12 and 24 dimensions are obtained.Then,through the Principal Components Analysis(PCA)to filter out the components that can effectively distinguish the percussion sound signal in different states of the slot wedge,the feature matrix after feature screening is used as the input for pattern recognition,which is conducive to identifying the corresponding state type of the percussion sound signal.(4)Stator slot wedge tightness classification identification.By using the Extreme Learning Machine(ELM)compared with the BP neural network,when entering the same eigenvectors,it can be observed that there are certain differences in the recognition accuracy of eigenvectors of different dimensions.Among them,the classification accuracy of the 24-dimensional Mel frequency cepstral coefficients matrix in the BP neural network reached98.33%.