Research On The Method Of Spindle Fault Prediction Based On Fan State Data

Fault prediction of key components of the fan is an important part of the intelligent wind power system.Effective and fast fault state prediction method is of great significance to improve the operation capability of the system platform,reduce the risk of failure and save maintenance costs.The data acquisition and monitoring(SCADA)system of the fan stores a large amount of data.How to realize the fault state prediction of the important components of the fan driven by the SCADA system data is a research hotspot in the current academic and industrial fields.Through the research and analysis of the current fault prediction methods in the field of wind power,it is found that the SCADA system of wind power usually collects only the instantaneous state data of the fan,and the fan features are single in the training modeling.Some algorithms lack the ability to represent the high-order features of wind turbines,which makes it difficult to mine the deep information of wind power system data.In addition,in the case of sparse fan data distribution,most machine learning models are difficult to effectively learn parameters,which is the main problem of fan component fault prediction.Aiming at the above problems,this thesis introduces different fault prediction schemes of fan spindle step by step from the point of view of characteristics,and puts forward the fault state prediction method of fan spindle based on sliding window feature and DCFM fusion structure successively.The main research contents are as follows:(1)In this thesis,a method for predicting the main shaft failure status based on the fan sliding window characteristics is presented,which is aimed at the single feature problem faced by building the model using the original characteristics of the fan,the method by constructing a series that can represent the current state of the fan characteristics,combined with the e Xtreme Gradient Boosting(XGBoost)algorithm in the field of classification of advantage,to eventually achieve the goal of ascension model classification effect.The single problem of fan features is solved by constructing sliding window features.Compared with the model trained with original fan features,the prediction effect of the model after integrating sliding window features is improved to a certain extent.(2)High and low order features of wind turbines contain rich state information of wind turbines.It is of positive significance to timely mining such information to improve the fault classification and prediction effect of wind power models.In this paper,Deep learning knowledge is introduced into the field of wind power fault prediction,and a neural Network structure integrating Factorization Machine(FM),Cross Network and Deep Network modules is constructed,which is named as DCFM.Compared with other deep learning methods,DCFM model is more detailed in learning low-order features,can also well learn parameters in sparse data space,and has certain ability to explore both explicit and implicit representations of high-order features.The comparative experiment on real data sets of wind farms shows that the DCFM model has good Area Under the Curve(AUC)classification performance index and Log-Likelihood Loss(Log Loss)convergence effect.Compared with other excellent models,it also has certain advantages and can meet the demand for fault prediction in the wind power field.