基于VME-M1DCNN-LSTM的齿轮异常状态智能识别

杜文友; 王宇琦; 崔霄; 徐伟; 崔建国

doi:10.3969/j.issn.2095-1248.2023.05.007

沈阳航空航天大学学报 ›› 2023, Vol. 40 ›› Issue (5) : 50 -55. DOI: 10.3969/j.issn.2095-1248.2023.05.007

信息科学与工程

基于VME-M1DCNN-LSTM的齿轮异常状态智能识别

杜文友 ¹ ,
王宇琦 ¹ ,
崔霄 ²^,³ ,
徐伟 ¹ ,
崔建国 ¹

作者信息 +

Intelligent recognition of gear abnormal states based on VME⁃ M1DCNN⁃LSTM

Wenyou DU ¹ ,
Yuqi WANG ¹ ,
Xiao CUI ²^,³ ,
Wei XU ¹ ,
Jianguo CUI ¹

Author information +

文章历史 +

PDF (896K)

摘要

针对工程实际中齿轮振动信号受噪声污染严重导致其异常状态难以准确识别的问题，提出了一种基于变分模态提取（variational mode extraction，VME）和多尺度一维卷积（multiscale one⁃dimensional convolution，M1DCNN）融合长短时记忆神经网络（long short⁃term memory，LSTM）的齿轮异常状态智能识别新方法。首先，采用VME方法分别对采集到的齿轮处于正常状态、轮齿碎裂、齿轮断齿、齿根裂纹以及齿轮磨损等5种状态的原始振动信号进行预处理，去除原始振动信号中的噪声干扰，提取齿轮不同状态的主模态分量作为齿轮状态的特征信息；其次，由提取的齿轮状态主模态分量构建训练数据集与测试数据集；最后，设计了M1DCNN-LSTM异常状态识别模型，并采用所构建的数据集对设计的异常状态识别模型进行了测试试验验证。结果表明，所提出的方法可以很好地实现齿轮异常状态智能识别效能，异常状态识别准确率达99.25%，明显高于其他相关齿轮异常状态识别方法。

Abstract

In engineering practice，the vibration signal of gears is severely polluted by noise，making it difficult to accurately identify their abnormal states. To address the problem，a new intelligent recognition method for gear abnormal states based on the variational mode extraction （VME） and the multiscale one-dimensional convolution （M1DC） fusion with long short term memory （LSTM） neural network was proposed. Firstly，the VME method was used to preprocess the original vibration signals in five states： normal state，gear tooth fragmentation，gear breakage，root crack，and gear wear. The noise in original vibration signals was removed，and the principal mode components of gears in different states were extracted as the feature information of the gear state. Secondly，a training data set and a test data set were constructed from the extracted principal mode components of the gear state. Finally，an M1DC-LSTM abnormal state recognition model was designed，and the constructed data set was used to test and verify the designed model. The results show that the method proposed in this paper can effectively achieve intelligent recognition of gear abnormal states，and the accuracy rate reaches 99.25%，which is significantly higher than other related methods.