Research On State Recognition Of Turbine Runner Blade Crack Based On Machine Learning Algorithm

As an important equipment for generating clean energy in hydropower generation,the failure of turbine has a serious impact on the project.As the core component of turbine,the blade crack fault of runner is a common fault type.The crack fault will affect the power generation efficiency and even affect the safe operation of the unit.The existing detection methods are mostly downtime detection or detection based on manual experience,and the detection efficiency is not high and the results are not accurate.Therefore,how to quickly detect the runner crack,get accurate judgment of crack fault through simple measurement and analysis,and prolong the operation life of the turbine unit is very meaningful.In this paper,the vibration signals of three crack states are measured by self-designed signal acquisition system through wireless transmission.The crack fault characteristics are extracted by signal preprocessing and variational mode decomposition.Then different machine learning algorithms are introduced to improve the algorithm and identify the crack fault.The main research work and achievements of this paper are as follows :(1)In this paper,a high-precision signal acquisition system is designed independently.The effectiveness of the acquisition system is verified by signal acquisition experiments.The vibration acceleration sensor is attached to the main shaft of the turbine,and the vibration signals of the blade without damage,micro-crack and large-crack state are measured by wireless transmission under rated conditions.(2)In order to deal with the noise influence of the measured signal during wireless transmission and unit operation,the wavelet packet denoising method is selected.By comparing the wavelet packet basis function and the number of decomposition layers,the optimal denoising parameters are selected for wavelet packet denoising.A simple time-frequency analysis is performed on the denoised signal.The change of the amplitude of the first three frequencies in the frequency domain diagram can determine whether the signal has a large crack signal,which has certain practical significance for the fault diagnosis of turbine cracks.(3)Through the improved WOA-VMD algorithm,the signal is decomposed,the decomposed components are screened and reconstructed,and then the characteristic parameters of the reconstructed signal are calculated to form the characteristic matrix.The characteristics of the fault are extracted to prepare for the subsequent pattern recognition,and the fault parameters are successfully extracted.(4)Introducing the BP neural network algorithm optimized by genetic algorithm,random forest algorithm and support vector machine algorithm are introduced to identify the fault pattern,and the recognition accuracy is improved by improving the support vector machine algorithm.The recognition accuracy of the algorithm for normal signal,micro-crack signal and large crack signal is increased,and the total recognition accuracy is increased from 93.13 % to 99.5 %.However,due to the complexity of the algorithm,the time consumption is increased.The recognition accuracy and calculation time of the machine learning algorithm used are compared to select the best recognition algorithm.By comparing the improved support vector machine algorithm,the recognition accuracy is the highest,up to 99.5 % and the recognition speed is the fastest,which meets the requirements of on-site crack recognition.