Power Equipment Fault Diagnosis And System Development Based On Deep Learning

The development of new energy represented by wind power generation has become the leading direction to solve the low-carbon energy transition.However,with the increasing installed capacity of wind power generation,the high cost of operation and maintenance has become a crucial problem.The gearbox itself,as one of the most expensive wind turbine equipment,accounts for a large proportion of the investment cost of the whole wind turbine.In order to minimize the economic loss caused by gearbox failure shutdown,it is very important to realize the monitoring technology of early warning signal for ensuring the safe and reliable operation of wind turbines.This paper takes wind power gearbox as the research object,based on massive historical operation data of SCADA system,proposes to build a datadriven model of gearbox under normal operation state based on feature extraction of deep autoencoder and long short-term memory neural network.Deep autoencoder is used to reduce the dimension of multi-source data in SCADA,and the extracted features are cascaded to LSTM neural network as input to realize the prediction of output time series data,which provides a model basis for the subsequent fault warning of wind turbines.Taking the distribution of abnormal data in the wind-power scatter plot as the entry point,the influence of different parameter combinations of local outlier factor algorithm and DBSCAN algorithm on the identification effect of abnormal data was analyzed,and an adaptive parameter selection DBSCAN algorithm was proposed to improve the detection effect of abnormal samples at the edge of normal working conditions.The preprocessed data were used to establish a DAELSTM gearbox prediction model,and the influence of different dimensionality reduction on the output accuracy of the model was discussed.Compared with a single LSTM model prediction method,the results showed that the prediction accuracy of the proposed method was significantly improved,and a more accurate data-driven model was provided for the effective early warning of gearbox faults.A fault threshold determination criterion based on the residual probability density distribution and confidence level analysis was established,and a 1.5MW wind turbine gearbox oil temperature overrun fault was taken as an example to verify.The sliding window method was combined to smooth the predicted absolute residual sequence to reduce the false positive rate.Through comparison,it is found that the data-driven method proposed in this paper can judge the existence of abnormal conditions earlier and more accurately than the single LSTM model on the premise of ensuring the occurrence of faults in advance,and gain more precious time for the timely maintenance of wind turbine equipment.Finally,a cloud service platform for wind turbine gearbox fault warning is developed based on cloud computing technology of API interface.A separate industrial APP is formed by secondary development of the main research results of this paper,and deployment experiments are carried out on the cloud service platform,so as to improve the utilization rate of computing resources,realize the sharing of algorithm functions and reduce the difficulty of users in using services.