| In the long-distance natural gas pipeline,the compressor unit driven by gas turbine bears the role of pipeline pressurization,which ensures the long-distance transportation of natural gas.Effective gas turbine real-time monitoring means can ensure production safety and prevent accidents.The model is an effective monitoring method.By comparing the prediction value of the model with the actual value,anomalies can be found immediately.Because of the complex structure of gas turbine,the mechanism model based on physical equation is sometimes difficult to establish,so the data-driven model based on machine learning has become a better auxiliary means.However,the training of data-driven model requires a lot of computing resources,which may lead to system delay and real-time degradation.This paper solves these problems by studying machine learning and distributed computing.Firstly,this paper establishes an accurate data-driven model.Based on three kinds of neural network structures,this paper establishes different models,optimizes the performance of each model through parameter configuration,and studies their sample dependence.Secondly,real-time online training is realized through distributed computing.In this paper,a distributed computing platform is built based on a master-slave computer cluster.The effects of communication topology,parameter aggregation period and data splitting algorithm on model training time,prediction accuracy and prediction stability are studied,and the optimal algorithm to meet the real-time requirements is given.Finally,a preliminary monitoring system is established.By building back-end computing platform and front-end interface,the data and instruction interaction between front-end and back-end are realized,and the availability of the platform is verified by a specific case.Finally,this paper achieves the goal of establishing a dual-axis gas turbine monitoring platform with high ease of use,real-time and accuracy. |
PDF Full Text Request