Research On Neural Network Of Wind Turbine Blade Health Monitoring

Wind energy is clean and renewable.In recent years,countries all over the world focus on the deployment of wind turbines to alleviate the energy shortage and environmental pollution.The wind turbine blade is a key component of wind power generation system.Its health state has a significant effect on the overall performance of the wind turbine.Wind turbine blades often work in complex natural environment,such as lightning,salt fog,rain and snow,and random changes of strong and weak winds,and face a variety of damage and challenges.If the damage of the blades cannot be found and repaired in time,it will cause great economic losses.Therefore,it is of great significance to establish a real-time online non-contact wind turbine blade health monitoring system for the normal operation of wind turbines.Although some wind turbine blade health monitoring technologies have been developed,these technologies still have their limitations and are difficult to deploy and apply in the field.Therefore,most wind power plants still adopt manual inspection methods.This dissertation carried out in-depth research on the detection algorithm of the health status of wind turbine blades,and innovatively proposed a wind turbine blade health monitoring system based on acoustic features and neural network to replace manual inspections,focusing on how to use and optimize neural network model algorithms to better identify and classify the fault types of wind turbine blades.The main content of the dissertation includes:(1)The construction and analysis of audio data set.We recorded audio of different wind turbines in the wind power plant,the measured audio data of the wind field are sorted out and classified based on the signal characteristics to construct the data set,and then the data set is visualized and analyzed in detail.(2)A wind turbine blade health monitoring system based on octave band and BP neural network is proposed.The system adopts a non-contact monitoring method based on acoustic signal,which can carry out real-time online monitoring of the health state of the wind turbine blades.The features of 1/6 octave frequency band energy were extracted from the audio data,and then sent to the BP neural network model for training and testing.The audio data were simulated at different speeds to expand the overall sample data.The accuracy of the model on the test set is increased by 1.4% compared with the original model.Finally,the recognition accuracy of the model was 95.27%,which verified that the model could effectively identify the normal and abnormal health states of wind turbine blades.(3)An algorithm based on log mel-spectrogram and deep learning model is proposed to monitor the health state of wind turbine blades.The abnormal samples are subdivided into 3categories according to defect features,and a data set of 4 categories is obtained by combining the normal samples.The log mel-spectrogram is used as the input feature,and the pre-trained weights of the ResNet,DenseNet and EfficentNet on ImageNet are used for migration learning,and the method of data augmentation and modification of the loss function is used to enhance the model.The original CNN model is improved and the CRNN model based on attention mechanism is proposed innovatively.The improved model pays more attention to the correlation in time domain and the connection between channels in the frequency domain.Compared with the original model,the accuracy is improved by 1.3%,and the inference speed is increased by 9%.The weighted F1 scores of the model with the best comprehensive effect on the test set reached 90.89%,and the inference speed was144 FPS,which proved that the model can quickly and effectively identify the four defect types of the wind turbine blades.