基于MOPSO-CNN模型的压缩机气阀故障诊断技术
Compressor valve fault diagnosis technology based on a MOPSO-CNN model
张平 1孙霖 2史建超 2李亚民3
作者信息
- 1. 国家管网集团联合管道有限责任公司西部分公司,乌鲁木齐 830011
- 2. 中海油能源发展装备技术有限公司,天津 300450
- 3. 唐山行世科技有限公司,唐山 063000
- 折叠
摘要
针对传统方法难以提取有效的气阀故障信号,无法建立气阀状态与信号间复杂映射关系的问题,将气阀振动信号转为频域信号输入卷积神经网络(CNN)进行气阀状态诊断,采用多目标粒子群算法(MOPSO)对CNN的超参数进行优化,构建自适应CNN模型,并针对分类结果进行可视化分析,探讨了不同训练测试比对分类准确率的影响.结果表明:MOPSO-CNN模型可完成数据降噪、特征提取和故障分类的一贯式处理,实现端到端的故障诊断,其分类准确率和训练时间均优于传统方法;通过t-分布随机邻域嵌入(t-distributed stochastic neighbor embedding,t-SNE)可视化分析,证明了CNN模型在逐层特征提取和特征分离上的优越性;所建立模型在不同训练测试比的条件下表现良好,对训练数据的需求量不大.研究结果可为往复式压缩机气阀故障诊断提供实际参考.
Abstract
It is difficult to extract effective valve fault signals by traditional methods,and the complex mapping re-lationship between valve states and signals cannot be established.In response to this problem,valve vibration sig-nals have been converted into frequency domain signals and input into convolutional neural networks(CNN).Multi-objective particle swarm optimization(MOPSO)was used to optimize the hyperparameters of CNN and an a-daptive CNN model was constructed in order to analyze the classification results visually.The influence of different training tests on the classification accuracy is discussed.The results show that MOPSO-CNN model can achieve consistent processing of data denoising,feature extraction and fault classification,and achieve end-to-end fault di-agnosis.The classification accuracy and training time of the MOPSO-CNN model are better than traditional meth-ods.The superiority of the CNN model in terms of feature extraction and feature separation was demonstrated using t-distributed stochastic neighbor embedding(t-SNE)visualization analysis.The model performs well under differ-ent training test ratios and requires little training data.These results provide a practical reference for reciprocating compressor valve fault diagnosis.
关键词
多目标粒子群算法(MOPSO)/卷积神经网络(CNN)/压缩机/气阀/故障诊断Key words
multi-objective particle swarm optimization(MOPSO)/convolutional neural networks(CNN)/com-pressor/valve/fault diagnosis引用本文复制引用
出版年
2024
NETL
NSTL
万方数据