基于卷积神经网络-长短时记忆神经网络的磨煤机故障预警  

Coal mill fault early warning technology based on CNN-LSTM network

作　　者：杨婷婷[1] 高乾 李浩千 吕游[1] 陈晓峰[2] YANG Tingting;GAO Qian;LI Haoqian;LYU You;CHEN Xiaofeng(School of Control and Computer Engineering,North China Electric Power University,Beijing 102206,China;North China Electric Power Research Institute Co.,Ltd.,Beijing 100045,China)

机构地区：[1]华北电力大学控制与计算机工程学院,北京102206 [2]华北电力科学研究院有限责任公司,北京100045

出　　处：《热力发电》2022年第10期122-129,共8页Thermal Power Generation

基　　金：国家重点研发计划项目(2021YFB2601405-4)。

摘　　要：为提高对磨煤机故障的事前预知能力,结合深度学习方法的优势,在传统长短时记忆(LSTM)神经网络的基础上,提出基于卷积神经网络-长短时记忆神经网络(CNN-LSTM)的磨煤机故障预警方法。选择与磨煤机堵煤故障相关的测点作为模型的输入量,进行多元时间序列预测。得到模型输出预测值与磨煤机正常工作状态下的运行数据之间的偏离度函数,运用核密度估计方法确定预警阈值,实现磨煤机堵煤故障预警。以某660 MW火电机组的中速磨煤机为研究对象,建立CNN-LSTM模型并进行故障预警试验。试验结果表明,该模型可以精确预测磨煤机多个测点参数的变化趋势,相较于LSTM神经网络模型具有更高的精确度。该方法能够提前对磨煤机堵煤故障做出有效预警。In order to improve the ability to predict the pulverizer fault in advance,combined with the advantages of deep learning method,a pulverizer fault early warning method based on convolutional neural network-long and short-term memory(CNN-LSTM)neural network is proposed on the basis of conventional LSTM network.The time series of coal blockage is selected as the input variables of multivariate prediction model.The deviation function between the predicted value of the model output and the operation data of the pulverizer under normal working state is obtained,and the early warning threshold is determined by using nuclear density estimation method to realize the early warning of coal blockage fault of the pulverizer.Taking the medium speed coal mill of a 660MW thermal power unit as the research object,the CNN-LSTM network model is established and the fault early warning experiment is carried out.The experimental results show that,this model can accurately predict the change trend of parameters at multiple measuring points of the coal mill,and has higher accuracy than the LSTM model.This method can give an effective early warning of coal blockage fault of pulverizer in advance.

关 键 词：磨煤机 故障预警 CNN LSTM神经网络 时间序列预测 

分 类 号：TK39[动力工程及工程热物理—热能工程]