Predictive Fault Warning And Diagnosis Of Complex Wind Turbine Operation

As widely application and fast development of wind energy among the globe,wind energy quantity demand is increasing year by year,the scale of wind farm and the size of wind turbine have been growth accordingly.Timely conditional monitoring and fault diagnosis are critical to reducing maintenance cost of wind turbine operation and supporting of electrical power system stability.With the rapid development of information technology and big data technology,data driven and artificial intelligence-based method have been fused with traditional quality monitoring and fault detection method in industrials.How to use above mentioned frame work to provide accurate failure warning and diagnosis and interpretable analysis for wind turbine operation is meaningful study.In this paper,using supervisory control and data acquisition datasets,basing on wind turbine conditional monitoring and fault diagnosis(WT-CMDF)which have based on traditional frequency pattern recognition models,we proposed new WT-CMDF which employed Bayesian deep learning models.For the computation complexity of traditional inference method for Bayesian deep learning model,it's not suitable to use for complex and large scale industrial datasets,therefore,in this paper we have adopted newly proposed method which combined Monte Carlo with dropout neural network to build new WT-CMDF system.The new system not only has improved on diagnosis accuracy but also provide interpretable uncertainty analysis and new diagnosis decision policy due to Bayesian characteristic which have been decreased the wrongly diagnosis risk and cost.First of all,we have shown background,significance and main direction of this research.Then we have summarized at large about related studies including CMDF of wind turbine subsystem and whole turbine machine aiming for different frequency signals,machine learning based WT-CMDF,main crucial model theory and inference method of Bayesian deep learning model,the flaw and disadvantage of present WT-CMDF have been analyzed.We then clarify creative WT-CMDF framework which combines uncertainty information and given executive summary and framing for this paper.Secondly,we have formulated general inference process of Bayesian machine leaning model and summarized commonly used gradient estimation method of variational inference objective function.We have also summarized theory foundation and derivation process for obtaining prediction uncertainty of Monte Carlo Dropout neural network which is as effective as Bayesian neural network.Thirdly,we have used experimental data to compare uncertainty quality of deep neural network and gaussian process model of difference Monte Carlo methods and different model structures.Then we go on with different compositions of prediction uncertainty and analyzed effect of combining heterogeneous aleatoric uncertainty into epistemic uncertainty for noisy input data.Fourthly,we have developed the new WT-CMFD of a practical wind turbine operation project based on Monte Carlo dropout neural network and gaussian process classifier summarized in section two.After we have finished the basic fault diagnosing and feature relevance analyzing task,we proposed we proposed “Reject” option for WT-CMFD using the obtained epistemic uncertainty and heterogenous uncertainty.The R-Accuracy metric which has combined “Reject” option into diagnosis cost function may reduce fault diagnosis risk.At last,based on finished research for deep Bayesian machine learning model and corresponding inference methods,we proposed Monte Carlo variational inference based deep categorical latent gaussian process model(DCLGP)for small scale multivariate categorical data.Through combining noise parameter with kernel function,we have reduced parameters of variational inference model.And it's effective to enhance correlation modeling of layer hidden variable by applying Monte Carlo reparameterization technique.The newly proposed model has improved some metrics performance compared with Monte Carlo dropout neural network in wind turbine fault diagnosis with uncertainty.