Application Of Time-frequency Analysis And Deep Learning To Identification Of Vortex Band In The Francis Turbine Draft Tube

Francis turbine is widely used in all kinds of hydropower station equipment due to its wide range of water head,small size,light structure and relatively simple design and maintenance.How to ensure its safe and stable operation has become a key industry issue in current research.The vortex will be formed in the draft tube when the unit deviates from the optimal operating conditions,leading to pressure pulsation,hydraulic vibration,unstable operation of the unit,which seriously threatens the safe operation of the whole unit because of the Francis turbine working principle and structural characteristics.Hence,it is necessary to conduct real-time condition monitoring and efficient analysis of the vortex.The thesis studies the visualization of the draft tube vortex identification through the Francis turbine draft tube vortex vibration signal by applying the time-frequency analysis method and convolutional neural network.The thesis determines the vibration signal characteristics of the draft tube under different operating conditions and discusses the application performance of different simultaneous frequency analysis methods in the visualization of the vortex of draft tube based on the measured vibration data of the draft tube by using convolutional neural network,hyperparameters such as different activation function,learning rate and gradient descent algorithm are selected to train convolutional neural network,to determine the influence of different hyperparameters on convolutional neural network to identify the draft tube vortex condition quickly and easily.The main contents and result as following:(1)The traditional single-time-domain or single-frequency-domain analysis methods cannot accurately show the frequency characteristics of pressure pulsation signals in the Francis turbine tailpipe over time.However,the variation characteristics of the vortex in different time periods should be considered,the thesis introduce the time-frequency analysis method that can consider both time-domain and frequency-domain characteristics to analyze the unsteady time-varying pressure pulsation signals generated by the vortex in the Francis turbine tailpipe in order to achieve the purpose of visualization of the vortex under different operating conditions by using time-frequency image.Compared with traditional signal analysis methods,time-frequency analysis can analyze the variation characteristics of pressure pulsation over time,which is more accurate and time-efficient in the field of vortex identification of draft tube.(2)Convolutional neural network,has good topology for geometric features of images.as a deep learning network.geometric transformation of features will not affect the process of feature extraction by convolutional neural network in the process of image recognition,the research establishes a convolutional neural network for image recognition of vortex of the Francis turbine and select different activation functions,learning rates,gradient descent algorithms and other super parameters training the convolutional neural network,to determine the influence of different super parameters on the convolutional neural network,and recognize and classify the time-frequency images of the Francis turbine under different vortex conditions.(3)The thesis finds the disadvantages and advantages of different time-frequency pressure pulsation signals and determines the characteristics of the vibration signal of the tailpipe under full operation conditions by combining time-frequency analysis methods to convert the vibration signal under different operation conditions into time-frequency image based on the measured vibration data of Francis turbine tailpipe of a hydropower station in Norway.The results show that the application of time-frequency analysis method and convolutional neural network to identify and classify the draft tube vortex has fast recognition speed,low training cost,and an accuracy of more than 98%,which has guiding significance in engineering practice.