摘要
本文针对冷水机组跨工况故障诊断过程中工况变化导致的模型性能下降问题,提出了基于领域对抗网络(DANN)的制冷剂泄漏故障跨工况诊断模型,基于迁移学习和对抗训练的思想将不同工况下的运行数据映射到一高维空间,在高维空间下将不同工况下的样本进行混淆,从而提高模型的泛化能力。本文从制冷剂泄漏故障检测和泄漏等级划分两个方面对比分析了DANN与一般模型在外延工况下的诊断准确率,结果表明不论是制冷剂泄漏故障检测还是泄漏等级划分,DANN在外延工况下均表现出了优于一般模型的诊断效果。在制冷剂泄漏检测问题中,DANN在部分工况下相较于深度神经网络(DNN)能够提升20%以上的准确率;在泄漏等级划分问题中,DANN的准确率在各个迁移方向上高于其他模型10.2%~27%。
In this paper, aiming at the problem of model performance degradation caused by working condition changes in the process of cross condition fault diagnosis of chillers, a cross condition diagnosis model of refrigerant leakage fault based on domain adversarial neural network(DANN) is proposed. Based on the idea of transfer learning and adversarial training, the operation data under different working conditions are mapped to a highdimensional space, so that samples under different conditions are confused in high-dimensional space, which improves the generalization ability of the model. The accuracy of DANN and general model is compared and analyzed from two aspects of refrigerant leakage fault detection and leakage level division. The results show that DANN has better diagnostic effect than general model in both refrigerant leakage fault detection and leakage level division. In the problem of refrigerant leakage detection, DANN can improve the accuracy by more than 20%compared with deep neural network(DNN) under some working conditions. And in the problem of leakage level division, the accuracy of DANN is 10.2% to 27% higher than other models in all migration directions.
作者
黄巍
赵博
曹承
朱旭
李鹏程
晋欣桥
HUANG Wei;ZHAO Bo;CAO Cheng;ZHU Xu;LI Pengcheng;JIN Xinqiao(Institute of Refrigeration and Cryogenics,Shanghai Jiao Tong University,Shanghai 200240,China;The Third Military Representative Office of the Naval Equipment Department in Shanghai,Shanghai 200093,China;Shanghai Marine Equipment Research Institute,Shanghai 200031,China)
出处
《制冷技术》
2022年第6期17-22,28,共7页
Chinese Journal of Refrigeration Technology
关键词
冷水机组
制冷剂泄漏
跨工况诊断
领域对抗网络
Chiller
Refrigerant Leakage
Cross-condition diagnosis
Domain adversarial neural network
