<p>The accurate and real-time fault diagnosis of reciprocating compressors is quite challenging, particularly in dynamic environments such as marine settings, where operating conditions change a lot over time. These variations, such as intake pressure, temperature, and gas component, can significantly alter the pressure-volume (<i>p</i>-<i>V</i>) diagram, leading to potential misinterpretations by traditional fault detection methods. This study proposes an intelligent framework that unites a novel isoparametric normalization of the indicator diagram with a lightweight convolutional neural network (CNN), making two contributions. First, the logarithmic <i>p</i>-<i>V</i> diagram is innovatively warped into a dimensionless square domain by normalization of isoparametric element mapping; this procedure magnifies fault-induced distortions while suppressing theoretical <i>p</i>-<i>V</i> diagram changes caused by load, speed, or gas component variations. Second, a compact CNN, systematically tuned through hyper-parameter optimization, is trained on the normalized diagrams. With the feature enhancement of isoparametric normalization, the network converges within three epochs (about 6 min) and attains 98.43% test accuracy on 960 labelled samples covering fifteen compressor states. This performance was validated through extensive experiments on a 160 kW two-stage test rig, where the model demonstrated robustness and reliability under varying operating conditions. The model, optimized for real-time deployment in dynamic industrial environments and designed to support continuous sample library enrichment and online retraining, offers a practical solution for predictive health management of reciprocating compressors.</p>

错误:搜索内容不能为空,请输入英文关键词
错误:关键词超出字数限制,请精简
高级检索

A novel isoparametric normalization mapping for p-V diagrams in intelligent diagnosis of reciprocating compressors via CNN

  • Qihang Li,
  • Nan Zhao,
  • Xiaomin Zhang,
  • Yanyu Cao,
  • Shubiao Hou,
  • Siqi Wu,
  • Weimin Wang

摘要

The accurate and real-time fault diagnosis of reciprocating compressors is quite challenging, particularly in dynamic environments such as marine settings, where operating conditions change a lot over time. These variations, such as intake pressure, temperature, and gas component, can significantly alter the pressure-volume (p-V) diagram, leading to potential misinterpretations by traditional fault detection methods. This study proposes an intelligent framework that unites a novel isoparametric normalization of the indicator diagram with a lightweight convolutional neural network (CNN), making two contributions. First, the logarithmic p-V diagram is innovatively warped into a dimensionless square domain by normalization of isoparametric element mapping; this procedure magnifies fault-induced distortions while suppressing theoretical p-V diagram changes caused by load, speed, or gas component variations. Second, a compact CNN, systematically tuned through hyper-parameter optimization, is trained on the normalized diagrams. With the feature enhancement of isoparametric normalization, the network converges within three epochs (about 6 min) and attains 98.43% test accuracy on 960 labelled samples covering fifteen compressor states. This performance was validated through extensive experiments on a 160 kW two-stage test rig, where the model demonstrated robustness and reliability under varying operating conditions. The model, optimized for real-time deployment in dynamic industrial environments and designed to support continuous sample library enrichment and online retraining, offers a practical solution for predictive health management of reciprocating compressors.