长短期记忆神经网络用于空调系统故障时间序列分析  

Fault Time Series Analysis for Air-conditioning System Based on Long-short Term Memory Neural Network

作　　者：宋海川 顾明伟 张弘韬 董小林 SONG Haichuan;GU Mingwei;ZHANG Hongtao;DONG Xiaolin(Gree Electric Appliances,Inc.of Zhuhai,Zhuhai 519070,Guangdong,China;Shanghai Nuclear Engineering Research and Design Institute,Shanghai 200030,China)

机构地区：[1]珠海格力电器股份有限公司,广东珠海519070 [2]上海核工程研究设计院有限公司,上海200030

出　　处：《制冷技术》2020年第6期24-30,共7页Chinese Journal of Refrigeration Technology

摘　　要：空调系统自动故障诊断已经成为保障空调机组安全稳定运行的重要手段。针对传统机器学习方法难以自学习和适应故障时间序列特征从而准确性下降的问题,结合长短期记忆(Long-short Term Memory,LSTM)神经网络适用于处理高度时间相关性和高维耦合性数据的特点,本文提出了一种基于LSTM的故障时间序列分析方法处理典型的故障前后时序数据,搭建故障智能诊断模型。采集实际运行的风冷螺杆机组低压保护故障时间序列数据,用于训练LSTM网络。结果表明:基于LSTM网络的模型在测试集上分类准确率达92.86%,验证了其相对于传统的机器学习算法具有更高的准确度,随着数据量的提升,LSTM有望能发挥其更好的预测性能。Automatic fault diagnosis has become an important means to maintain air-conditioning systems steadily and safely running with modules and equipment.Aiming at problem of decreasing performance in that the traditional machine learning methods are often difficult to self-study and be adapted to the characteristics of fault time series,considering the characteristics that long-short term memory(LSTM)network is fit to process high-dimensional,strong coupling and high time-dependent data,this paper proposes a fault time series analysis method based on LSTM to deal with the typical time-series data before and after the fault moment and builds fault intelligence diagnostic model.The fault time series data of low-pressure protection of air-cooled screw units are collected and experimented to train LSTM neural network.The results show that LSTM-based network scores 92.86%on the test sets and scores more than that using traditional machine learning.With the increase of data volume,LSTM is expected to exploit the advantages in time sequences prediction performance to the full.

关 键 词：长短期记忆网络 深度学习 时间序列分析 故障诊断 空调系统 

分 类 号：TU831[建筑科学—供热、供燃气、通风及空调工程] TP306.3[自动化与计算机技术—计算机系统结构]