Research On Vibration Fault Diagnosis Of Hydropower Units Based On Neural Network

As one of the clean energy sources,hydropower plays an important role in China’s energy development.In order to effectively ensure the stable operation of hydropower units,real-time monitoring and analysis of equipment operating conditions is required.According to statistics,80% of the fault characteristics of hydropower units can be reflected in the vibration signals of the unit,so it is extremely important to monitor and analyze the vibration signals of the hydropower units.When using the vibration signals of the hydroelectric generating unit to perform state analysis and fault diagnosis of the equipment,etc.,the characteristics of the abnormal vibration signals need to be acquired first,that is,the characteristics of the vibration signals are extracted.At present,the vibration signal feature extraction generally includes methods such as time domain,frequency domain,and time-frequency domain analysis.In this paper,the wavelet packet decomposition method is selected from the perspective of time-frequency analysis to extract the energy characteristics of equipment vibration signals,which lays the foundation for the next equipment fault diagnosis.However,in the actual production process,noise may affect the recognition rate of fault diagnosis and lead to misdiagnosis of equipment faults.In order to solve this problem,this paper uses wavelet noise reduction method to reduce noise of noisy signals.However,because the selection of the threshold function in the wavelet denoising process directly affects the noise reduction effect,which leads to a certain degree of defects in the traditional wavelet threshold noise reduction,this paper improves the wavelet threshold function and tests it with simulation signals.It can be seen from the signal-to-noise ratio and the root-mean-square error index that the noise reduction effect of the reconstructed signal is good.After de-noising vibration signals and extracting energy feature vectors,it is necessary to identify and diagnose equipment fault types through fault diagnosis models,so an accurate and stable fault diagnosis model needs to be established.Because the long-term and short-term memory neural network has great advantages in dealing with non-linear mapping and the problem of seriality of data,it is very suitable for vibration fault diagnosis of hydraulic turbine units(it has serious coupling relationships between components of hydroelectric equipment,vibration signals and faults There is a non-linear relationship between the types,and the vibration signals extracted by energy features have certain sequence characteristics.Therefore,a long-term and short-term memory neural network is used to establish a fault diagnosis model.Collecting data and conducting simulation tests,and analyzing from the training and test set accuracy and network convergence rate,the model has achieved good results.However,during the establishment of a long-and short-term memory neural network-based vibration fault diagnosis model for hydraulic turbine units,the random selection of hyperparameters has a large impact on the accuracy of neural network recognition.The number of nodes and the inhibition rate of the neural network(its role is to enhance the stability of the network)are optimized to make the trained model reach a better state.After simulation testing and comparative analysis,the parameter-optimized fault diagnosis model simplifies the neural network model while ensuring recognition accuracy,the convergence rate is improved,and the model achieves the expected results.