Research On Condition Monitoring Of Steam Turbine Based On Data Driven

Steam turbine is the core equipment of power generation system.If its abnormal operating conditions can’t be found and handled timely,there will cause serious faults and safety accidents.So it’s important to find the abnormal conditions of steam turbine to ensure safety in production by monitoring its operating conditions.This paper builds the condition monitoring model of steam turbine by researching the multivariate time series produced in the process of steam turbine operating.The main works are as follows:（1）The time sequence data of steam turbine is preprocessed and a denoising algorithm based on the EEMD combining with Wavelet transform is proposed.After dealing with the missing values and outlier,noise is removed by EEMD-Wavelet transform.First,every variable in different states was decomposed by EEMD algorithm.The IMFs which has low correlation with original signal was removed.Next,every retained IMF was decomposed to a group of wavelet coefficients and denoised by wavelet threshold.Finally,the signal is reconstructed.The experimental results show that the EEMD-Wavelet transform algorithm is better than single EEMD algorithm or Wavelet transform algorithm.（2）An 1D-CNN model of condition monitoring for steam turbine based on actual data is built.First,making samples of turbine states by multiple sliding windows and dividing them into training set and testing set according to the proportion of 60% and40% to train and test the model.Next,the 1D-CNN network is built by the framework of Tensorflow and its initial parameters is set.Then,the optimizer,learning rate,batch size and epoch are chosen by training the model many times.Finally,the experimental results show that the 1D-CNN network has high testing accuracy about the different conditions of steam turbine.（3）Condition monitoring model of steam turbine based on the improved CNNGRU network is proposed.First,the model of CNN-GRU has the ability of extracting spatial features and timing features by combing CNN and GRU.Next,the model is improved by using new activation function,replacing the full connection layer to global average pooling layer and using self-ATTENTION to highlight the key features and avoid over fitting.The experimental results show that the improved CNN-GRU model has good generalization ability and its multiple indicators are better than 1DCNN model.Comparing with others popular methods,the CNN-GRU network has the best performance.