基于CNN-LSTM的燃气轮机燃烧室故障预警  

Fault Warning of Gas Turbine Combustor based on CNN-LSTM

作　　者：周锐 康英伟 ZHOU Rui;KANG Ying-wei(School of Automation Engineering,Shanghai University of Electric Power,Shanghai,China,Post Code:200090)

机构地区：[1]上海电力大学自动化工程学院,上海200090

出　　处：《热能动力工程》2024年第1期191-197,215,共8页Journal of Engineering for Thermal Energy and Power

基　　金：国家自然科学基金项目(61573239);上海发电过程智能管控工程技术研究中心资助项目(14DZ2251100)。

摘　　要：为了解决燃气轮机燃烧室中的故障问题,结合深度学习的优势,在长短期记忆网络(Long Short-term Memory,LSTM)的基础上,提出了一种基于卷积神经网络-长短期记忆网络的燃烧室故障预警方法。首先,根据正常的历史运行数据构建燃烧室的预警模型,再将特征参数输入到预警模型中得到预测值,预测值与实际值之间的偏离度可以反映燃烧室内部工作是否正常。考虑到模型预测结果的非平稳性和非线性等特点,引入滑动窗口法确定故障预警阈值,最后根据确定的预警阈值判断是否出现故障。采用某燃气-蒸汽联合循环发电机组仿真平台对上述模型进行验证。仿真结果表明:该模型相较于LSTM预测模型具有更高的精确度,可以及时发现故障征兆,并对燃烧室故障做出有效预警。In order to solve the fault problem of gas turbine combustor,a combustor fault early warning method based on convolutional neural network(CNN) and long short-term memory(LSTM) network was proposed combining with the advantage of deep learning.First,a prediction model of the combustor was constructed based on the normal historical operation data.Then,the characteristic parameters were input into the early warning model to obtain the predicted values.The deviation between the predicted and actual values could reflect whether the internal work of the combustor was normal or not,and considering the nonstationary and nonlinear characteristics of the model prediction results,the sliding window method was introduced to determine the fault warning threshold.Finally,whether a fault occurs was judged according to the determined warning threshold.The above model was validated on a gas-steam combined cycle generator unit simulation platform.The simulation results show that the model has higher accuracy than the LSTM prediction model,and can detect the signs of failure in time and make effective early warning of the combustor failure.

关 键 词：燃烧室 故障预警 LSTM神经网络 卷积神经网络 预测偏离度 滑动窗口法 

分 类 号：TK472[动力工程及工程热物理—动力机械及工程]