Fault Diagnosis Of Wind Turbine Drive Chain Based On Convolutional Neural Network

With the proposal and steady progress of the national "3060" plan,wind power generation,as an important component of new energy generation,has steadily increased its annual installed capacity.The status detection and fault diagnosis of wind turbine mechanical equipment is of great engineering significance.This article analyzes the possible faults that may occur in the gearbox of wind turbines,and selects bearings and gears in the transmission chain as typical fault units for research.Obtain the original time-domain information of the signal through vibration sensors.Using traditional signal analysis methods for fault analysis of signals,using complex Morlet wavelets and Grami angle fields to extract features from the data,in order to obtain the implicit internal connections of the signal.The vibration signal of the bearing data of Case Western Reserve University is subject to Fourier transform(FFT)to obtain the frequency domain information,which is used as the input data to build a onedimensional convolutional neural Network tomography diagnosis model.By adding a customizable convolution kernel,the early information filtering ability of the network is improved,and the final convergence speed of the network is improved.Verify the recognition ability of one-dimensional convolutional neural networks in frequency domain data classification tasks by designing six classification tasks;By comparing and verifying the performance superiority of one-dimensional convolutional neural networks with definable convolutional kernels through designing very similar tasks.Compared with BP(Back Propagation)neural network,1D-CNN and LSTM(Long Short Term Memory)neural network,1D-CNN network with definable convolution kernel has higher precision and faster convergence speed under the same model parameters.Considering that it is difficult to obtain effective fault signals on the diagnosis side in normal engineering practice,which affects the data richness of the model,resulting in a long training cycle,this paper proposes a transfer learning model based on complex Morlet wavelet and Gramian angular field(GAF)and other feature extraction methods,which fine tunes the model structure on the basis of residual neural network(Resnet),Improved classification ability.In cross task transfer diagnosis tasks,complex Morelet wavelets are used to extract features from the original data and serve as network inputs to obtain model comparisons in multi task situations.In the cross model and cross working condition task,GEZ gear data and Yilan turbine gearbox data were selected as datasets,and Grami and angle fields were used to extract features from the data to explore the performance of the model in the cross model and cross working condition task.After comparison with the 2D CNN model,the improved Resnet Model classification ability was verified under the same experimental conditions.