Research On Rotor Unbalance Fault Identification Method Based On Convolutional Neural Network

Wind farms are typically built in outlying and harsh environments,and wind turbines inevitably suffer from blade ice and pitch angle deviations during operation,if detect and dealt with timely,can prevent life time and economic losses.Therefore,this thesis concentrates on rotor imbalance fault detection,and it develops a rotor unbalance fault detection model integrating nacelle vibration signal time-frequency data and a convolutional neural network.The details are as follows:Built on the concept of wind turbine operations,a model of a 3MW wind turbine was developed in GH-Bladed simulation software.Two time-frequency analysis methods,short-time Fourier transform and wavelet transform,are utilized to compare and evaluate the impact of these two ways on the expression of the attributes and to select the wavelet time-frequency transform method to provide a basis for subsequent signal characterisation.The nacelle vibration signal features of rotor imbalance are investigated,and unbalance fault conditions are defined in the model,such as blade mass unbalance of 2% and aerodynamic unbalance of 4°,and time domain,frequency domain and time-frequency domain signal analysis methods are utilized.To just get the time-domain variation feature,the simulation results of the situations are compared with the nacelle vibration signals under normal conditions in the time domain.To acquire the frequency domain characteristics of the flawed signal,the wavelet energy spectrum analysis method is utilized in the frequency domain to contrast the change in the energy portion of the first frequency band of the normal signal with that of the faulty signal.The wavelet transform method is utilized in the time-frequency domain.This algorithm includes time-domain and frequency-domain fault characteristics to provide a more complete representation of signal properties,therefore,the data set is created based on the time-frequency map of the cabin vibration signal.The detection of rotor imbalance faults is discussed.To evaluate the influences of various convolutional layers,convolutional kernel sizes,and activation functions on the model,a convolutional neural network model based on 4-layer convolutional layers,3 × 3-sized convolutional kernels,and the Relu activation function is produced.The model is optimised by improving the size of the convolutional kernel and the activation function,and the study indicated that the precision of the optimised model is increased by 2%.Finally,the unimproved and improved models are tested with samples of normal operating conditions without noise reduction,samples of pneumatic imbalance conditions and samples of mass imbalance conditions.The results show that the optimised model has the best recognition capability for multiple conditions.