Research On Gas Path Fautl Diagnosis In Gas Turbine Groups

Gas turbines are widely used in the industrial field,and their structures are complex,the operating environment is harsh,and they are prone to failure.Effective fault diagnosis technology is of great significance to ensure the safe operation of the unit and reduce the maintenance cost in the life cycle.Most of the existing gas path fault diagnosis is applied on a single gas turbine.It is impossible to establish an accurate model of a new commissioning unit because of the lack of historical data.For gas turbines that have been running for a long time,the accumulated large amount of historical data provides a basis for data analysis of new gas turbines.However,the same type of gas turbine has large individual differences due to manufacturing errors a nd different maintenance time.It is impossible to directly use historical data for different gas turbines.Therefore,there is an urgent need for a new method that can effectively utilize the experience and knowledge contained in historical data to help t he target task establish a model and further improve the accuracy of gas path diagnosis.This paper hopes to solve the problem of the incomplete gas path fault diagnosis of the new gas turbine with the label data and the single gas turbine fault sample by means of transfer learning.The main work and innovations are as follows:(1)For the first time,a new research direction of gas turbine fault diagnosis is proposed.The definition of fault diagnosis of gas turbine groups is given,and a simulation model of gas turbine groups that can correctly reflect gas path fault is established.The data is obtained through simulation analysis,and the translatable knowledge of typical gas path faults under variable working conditions is extracted.The Finetune transfer learning method suitable for gas turbine group fault diagnosis is proposed through comparison and identification.(2)For the scenario where the new gas turbines have less marked data,the gas turbine with rich operational data is used as the source doma in,and the pre-trained model is established based on the convolutional neural network(CNN)to extract low-level fault features.Common knowledge(changes in health parameters)that are not affected by operating conditions and environmental changes at the time of failure is extracted between different gas turbines.After using the Finetune method to transfer the shared knowledge to the target domain,the proposed method can achieve accurate fault classification in the new gas turbines,which is more effective than the traditional data-driven methods.The importance of transfer decreases as the amount of data in the target domain increases.By analyzing the inherent patterns of each category by layer-by-layer visualization of the network,the rationality of common knowledge extraction is revealed.In addition,it is concluded from the experiment that as the data distribution difference of different gas turbines increases,the number of layers that needs to be tuned increases.(3)This paper uses the Finetune model to solve the problem of incomplete fault samples by integrating the multi-source domain fault samples.Through experiments,this paper establishes a model for the target domain category and the source domain from the same type of gas turbines and different type of gas turbines,which helps the classification task in target domain achieve better performance.It's verified that transfer learning has good generalization performance.(4)Finally,the recurrent neural network(RNN)model for the fault diagnosis of gas turbines under dynamic working conditions is proposed.The model that considers time characteristics is superior to the other models,and the simple mixture of data obtains poor diagnosis performance.This paper provides theoretical guidance for the practical applications of transfer learning in the process of gas turbine fault diagnosis through the systematical study of different actual scenes.