张利军,段礼祥,万 夫,张 健,刘香玉.往复压缩机故障的残差网络诊断方法[J].电子测量与仪器学报,2021,35(5):38-46 |
往复压缩机故障的残差网络诊断方法 |
Fault diagnosis method of reciprocating compressor based on residual network |
|
DOI: |
中文关键词: 往复压缩机 故障诊断 残差网络 ResNet50 |
英文关键词:reciprocating compressor fault diagnosis residual network ResNet50 |
基金项目:国家自然科学基金(51674277)、中石油战略合作科技专项(ZLZX2020 05 02)、中国石油集团公司项目(2019 F30)资助 |
|
|
摘要点击次数: 692 |
全文下载次数: 4 |
中文摘要: |
往复压缩机结构复杂、激励源众多,极易发生故障。 由于故障特征设计困难,且多依靠经验,导致传统方法诊断能力不
强。 基于卷积神经网络(convolutional neural networks,CNN)的智能诊断方法无需提取特征,可实现端到端的故障诊断,但存在
提取故障特征不准确、模型参数量大、训练时间长等难题。 为此,提出基于 PyTorch 深度学习框架的往复压缩机故障诊断方法
MPMRNet(multiple-processes-mini-ResNet)。 该方法采用多进程加载数据,以残差网络 ResNet50 为基础网络框架进行深度和宽
度的缩减,Adam 优化网络、StepLR 策略调整学习率,自动处理振动信号时频图像并进行敏感特征深度挖掘和评估。 通过多组
实验对比,该方法明显缩短了模型训练时间,权重参数量由 94. 1 缩小到 0. 58 M,模型复杂度由 4. 11 下降到 0. 21 G,显存占用率
由 37. 08%下降到 10. 92%,故障诊断的准确率达到 98. 28%,模型的诊断能力得到了明显提高。 |
英文摘要: |
The failure of reciprocating compressor happens frequently for the complex structure and rich excitation sources. Due to the
difficulty of designing fault features and relying on experience, the traditional methods have no strong diagnosis ability. The intelligent
diagnosis method based on convolutional neural networks ( CNN) can realize end-to-end fault diagnosis without feature extraction.
However, there are some problems such as inaccurate extraction of fault features, large number of model parameters and long training
time. Therefore, a fault diagnosis method of reciprocating compressor based on PyTorch deep learning framework MPMRNet (multipleprocesses-mini-ResNet) is proposed. In this method, multiple processes are used to load data, ResNet50 is taken as the basic network
framework and its depth and width are reduced. Adam and StepLR strategy are used to optimize the network and adjust the learning rate,
respectively. And time-frequency images of vibration signals are processed automatically to deeply mine and evaluate sensitive features.
Multiple comparison experiments show that this method significantly shortens the training time of the model, reduces the number of model
weight parameters from 94. 1 to 0. 58 M, the complexity of the model from 4. 11 to 0. 21 G, and the memory occupancy rate from
37. 08% to 10. 92%, and the fault diagnosis accuracy is up to 98. 28%, the diagnostic ability of the model is obviously improved. |
查看全文 查看/发表评论 下载PDF阅读器 |
|
|
|