Power Plant Fan Fault Early Warning Based On Data Driven

With the further implementation of the "double carbon" target,China’s thermal power generation has entered the era of "large units",which has further increased the operational requirements for power plant equipment.The induced draft fan has complex working conditions and harsh working environment.It is the equipment with the highest failure rate among fan equipment.The unscheduled downtime caused by fan failure will bring economic loss to the power station.The massive historical operation data of the equipment obtained by the data acquisition system of the power plant still needs further mining and development.In this paper,through a data-driven modeling approach,a neural network-based bearing vibration signal monitoring model and a multivariate monitoring model based on nonlinear estimation techniques are established for real-time monitoring of wind turbine status,and finally,early warning of faults is realized by means of probabilistic and interval statistics.The main works are:Studied the principle of induced draft fan and its internal mechanism,conducted statistics on various types of typical faults of induced draft fan,and summarized the causes of fan faults and various types of fault characterization signals.Introduction of monitoring variables for fan operation,pre-processing of data and preparation of modeling.A composite model based on the fusion of convolutional neural network and gated recurrent unit is studied.The vibration signal of fan bearing is used as monitoring variable to monitor and warn the fan condition;for the situation that there are many measurement points and mutual coupling,a feature selection method based on Max-Relevance and Min-Redundancy which computed by distance correlation coefficient is proposed to screen out the strongly correlated feature variables and eliminate the redundant ones.In order to solve the difficulties of manual setting of hyperparameters in neural networks and improve the prediction accuracy of the model,a Bayesian optimization algorithm is used for hyperparameter optimization;after simulation and comparison experiments,the prediction model has higher accuracy and realizes real-time monitoring;an early warning method based on the combination of Bayes test and sliding window is proposed,which can fully exploit the a priori information and obtain the residual mean and variance through moment estimation,while quantifying the failure probability.It is experimentally verified that this method is more timely and effective than the ordinary sliding window warning while keeping the false alarm rate and and leakage rate very low.A non-parametric multivariate state estimation technique model for induced draft fan multivariate state monitoring method is studied,and an improved multivariate state estimation technique model for dynamic memory matrix update is proposed,firstly,the density peak algorithm is used to cluster the historical samples to obtain the alternate memory matrix,and the dynamic memory matrix is selected in real time according to the similarity distance between the historical samples and the observation vector for calculating the alternate memory matrix,in contrast to the traditional memory In contrast to the traditional memory matrix construction method,the number of historical samples involved in the real-time calculation is relatively reduced by 2/3,and the prediction accuracy of the monitored variables is significantly improved.For the multivariate monitoring model,a sliding window based adaptive threshold similarity warning method is constructed,which simplifies the computational complexity compared with Bayes test and achieves fault warning.