| Turbine rotor is a key component of power plant equipment,once the failure will cause incalculable economic loss or even casualties,so it is very important to find a fast and accurate fault identification method.Past fault identification process,not only includes signal acquisition,signal processing and other professional knowledge,also need to mechanical running state and fault related background knowledge to understand somewhat,increasing the difficulty of fault diagnosis,design time consuming and difficult to ensure commonality,therefore to discuss how to simplify the process of fault diagnosis,at the same time improve the fault diagnosis accuracy and antinoise ability,is a great research value and practical significance of the research work.In recent years,as artificial intelligence has made continuous breakthroughs in many fields,deep learning has been widely applied.The powerful computing power of computers can learn the potential rules under big data,which makes signal processing more convenient.Taking the rotor of the turbine as the research object,an Improved Deep Convolutional Neural network(IDCNN)is proposed in this paper,which can identify and analyze the common faults of the rotor of the turbine accurately and efficiently.The main research work and results are as follows:Based on the analysis of fault monitoring technology and fault vibration mechanism of steam turbine rotor,the present laboratory equipment is used to build a turbine rotor fault simulation test bench.Simulation experiments are carried out for the four states of normal operation of turbine rotor,misalignment of rotor,unbalance of rotor and dynamic and static colliding and grinding,fault vibration signals are collected,and the original signals are pretreated,laying a foundation for the next step of signal analysis.Combined with the basic theory of neural network and the basic idea of deep learning,a one-dimensional convolutional neural network model for rotor fault diagnosis of steam turbine is constructed.The model is trained and tested by using the fault data collected from the simulation experiment.The comparison results show that the method is effective.Thus,it provides a strong support for further improving the performance of the model by improving the method of convolutional neural network.An improved framework model of deep convolutional neural network for turbine rotor fault diagnosis is proposed.The first convolutional layer of the network model uses a large convolution kernel.It is proposed to add failure input before the first convolutional layer to effectively improve the model’s anti-noise performance.Further,batch normalization processing after each layer of convolution operation is proposed.This model reduces the design difficulty of convolutional neural network and has a fast convergence speed,which greatly improves the identification efficiency of the model.Especially in the case of big data,the identification accuracy is relatively high.In addition,the global average pooling layer is proposed to replace the full connection layer in the model,which greatly reduces the number of parameters and the risk of overfitting in the traditional convolutional neural network model.Based on statistical indicators,through the simulated fault data for performance analysis of the model,Classification and Regression Tree algorithm(CART),K Nearest Neighbor algorithm(KNN),Support Vector Machine(SVM)and LeNet-5 comparing several traditional algorithms,and verify the identification capability of the presented model noise resistance.The results show that the recognition accuracy of CART and KNN algorithm is not more than 70%,the recognition accuracy of SVM algorithm is 82%,and the recognition accuracy of IDCNN diagnostic model proposed in this paper can reach 95%.IDCNN fault diagnosis algorithm has higher recognition accuracy and higher stability than SVM and Lenet-5 classical algorithms under different noise test sets,and the accuracy can reach more than 88%under different SNR,showing great advantages. |
PDF Full Text Request