Research On Fault Diagnosis For Hydropower Unit Based On Time-frequency Diagram And Convolutional Neural Network

Hydropower unit is the core equipment of a hydropower station,with its status related to the safe and stable operation of the hydropower station and even the whole power grid.The vibration signal contains the complete state information of the unit,but its influencing factors are numerous and coupled with one another.This makes the extraction of signal characteristics and the identification of fault types very difficult.The application of advanced achievements in signal processing and artificial intelligence,such as time-frequency analysis and convolutional neural network,a deep learning algorithm,is of great significance to improving the accuracy of fault diagnosis and the safety of hydropower units.The main contents of this paper are as follows:Firstly,the vibration mechanism of hydropower units and the time-frequency diagram of vibration signals were studied.In this paper,the vibration faults of hydropower units were classified;the causes,manifestations and influencing factors of various faults were analyzed.And the characteristics and characteristic frequencies of common vibration faults were summarized.Time-frequency analysis methods such as short-time Fourier transform,wavelet transform and Hilbert Huang transform were studied.Considering different window lengths and wavelet basis functions,time-frequency diagrams of typical simulation signals were constructed respectively.The feature expression performances of various methods were compared and analyzed.The results show that the time-frequency diagrams constructed by complex Morlet wavelet transform have the best time-frequency resolution.Secondly,the convolutional neural network model was studied.The basic principle and structure of convolutional neural network was introduced.The training process was analyzed.The commonly used optimization algorithms were compared.The causes and solutions of over fitting were discussed,and the classification process of convolutional neural network was introduced.As one of the important algorithms of deep learning,convolutional neural network contains a feature extractor,and can directly input two-dimensional original data,which has excellent classification ability for image recognition.Therefore,the time-frequency diagram of hydropower unit's vibration signals can be used as the input of the network.Finally,a diagnosis model based on time-frequency diagram and convolutional neural network was established to diagnose the actual faults of the hydropower unit.To start with,the vibration data of two identical units in a hydropower station were selected for diagnosis.In the process of diagnosis,the number of convolution kernels,the size of batches and the number of iterations were selected,considering the classification accuracy and diagnosis time.The diagnosis results show that the convolutional neural network has a good recognition rate for time-frequency diagrams.By comparing three time-frequency diagrams,it is found that the recognition rate of time-frequency diagram based on wavelet transform is higher.Then,the vibration data of different hydropower units were selected for diagnosis,and a high classification accuracy was achieved,which verifies the effectiveness and accuracy of the proposed method for fault diagnosis of hydropower units,and its good generalization performance.