基于深度极限学习机的燃气轮机气路故障诊断技术研究  

Gas path fault diagnosis technology of a gas turbine based on a deep extreme learning machine

作　　者：宋文强 沈登海 SONG WenQiang;SHEN DengHai(Western Pipeline Co.,Ltd.,National Pipe Network Group,Urumqi 830013,China)

机构地区：[1]国家管网集团西部管道有限责任公司,乌鲁木齐830013

出　　处：《北京化工大学学报（自然科学版）》2024年第4期89-98,共10页Journal of Beijing University of Chemical Technology(Natural Science Edition)

基　　金：国家管网集团重点科研项目(CLZB202204/DTXNY202204)。

摘　　要：为提高输气管道设备中燃气轮机的可靠性和可用性,在稳态仿真模型的基础上,形成包含现场可测量参数、健康状态参数和故障类型的故障数据库。从参数动态调整和空气质点混沌初始化两个方面对风驱动(wind driven optimization, WDO)算法进行改进,再利用改进算法(improved wind driven optimization, IWDO)对深度极限学习机(deep extreme learning machine, DELM)的超参数进行寻优,并试算不同模型结构对分类效果的影响,最终形成最优IWDO-DELM组合模型。结果表明,仿真模型的热力和水力参数准确,可以为故障数据库的生成提供基础;最优DELM的模型结构为9-81-44-1,激活函数均为Sine;IWDO-DELM模型在训练集和测试集上的故障分类准确率分别为99.12%、98.83%,优于支持向量机(support vector machine, SVM)、反向传播神经网络(back propagation, BP)、相关向量机(relevance vector machine, RVM)和极限学习机(extreme learning machine, ELM)等模型的计算结果。通过现场验证,证明了IWDO-DELM模型可有效识别燃气轮机气路上的单故障和多故障类型。研究结果可为输气管道的安全平稳运行提供实际参考。To improve the reliability and availability of gas turbines in gas pipelines, a fault database including field measurable parameters, health state parameters and fault types was formed on the basis of the steady-state simulation model. The wind driven optimization(WDO) algorithm was improved(IWDO) from two aspects of dynamic parameter adjustment and chaotic initialization of air particles, and the superparameters of the deep extreme learming machine(DELM) model were optimized by the improved algorithm. The influence of different model structures on the classification effect was calculated, and finally the optimal IWDO-DELM combination model was formulated. The results show that the thermal and hydraulic parameters of the simulation model are accurate, and can provide a foundation for the generation of fault database. The model structure of the optimal DELM was 9-81-44-1, and all activation functions were Sine. The fault classification accuracy of IWDO-DELM model using the training set and a test set was 99.12% and 98.83%, respectively, which is superior to the results using support vector machine(SVM), back propagation(BP), relevance vector machine(RVM) or extreme learning machine(ELM) models. Through field verification, it is proved that the IWDO-DELM model can effectively identify single and multiple fault types of a gas turbine gas path. The research results can provide a practical reference for the safe and stable operation of gas pipelines.

关 键 词：深度极限学习机 风驱动算法 燃气轮机 气路故障 分类准确率 

分 类 号：TE974[石油与天然气工程—石油机械设备]