| 邱大伟* **,刘子辰*,周一青* **,龙隆* **,谭雯雯***,曹欢* **.基于Transformer神经网络的滚动轴承故障类型识别[J].高技术通讯(中文),2021,31(1):1~12 |
| 基于Transformer神经网络的滚动轴承故障类型识别 |
|
| |
| DOI：10.3772/j.issn.1002-0470.2021.01.001 |
| 中文关键词: 滚动轴承； 故障类型识别； Transformer神经网络； 前向特征矩阵； 后向特征矩阵； 归一化位置编码； 权重增强 |
| 英文关键词: rolling bearing, fault type detection, Transformer neural network, forward feature matrix, backwardfeature matrix, normalization positional encoding, weight enhancement |
| 基金项目: |
| 作者 | 单位 | | 邱大伟* ** | | | 刘子辰* | | | 周一青* ** | | | 龙隆* ** | | | 谭雯雯*** | | | 曹欢* ** | |
|
| 摘要点击次数: 3041 |
| 全文下载次数: 1769 |
| 中文摘要: |
| 工程应用中的滚动轴承故障类型识别要求同时具有较高的识别准确度和时间效率，基于上述需求提出基于Transformer神经网络的滚动轴承故障类型识别方法。所提方法结合小波包变换时频域能量特征和快速傅里叶变换频域特征生成满足Transformer神经网络的输入样本矩阵，解决Transformer神经网络的输入问题。同时，提出应用于滚动轴承故障类型识别的归一化位置编码方法，解决Transformer神经网络在滚动轴承故障分析领域的位置编码问题。在此基础上，提出Transformer神经网络双向输入样本矩阵处理机制和算法训练过程中错误样本权重增强机制，提升所提方法的鲁棒性。使用KAt数据中心的滚动轴承数据集验证所提方法的识别性能，与现有常用深度学习方法相比，所提方法在时间效率和准确度性能上均有一定的优势，其中，准确度能够提升11%以上，单个样本的平均处理时间小于1ms。 |
| 英文摘要: |
| Rolling bearing fault type detection under practical applications requires both high detection accuracy and high time efficiency. Based on the above requirements, a rolling bearing fault type detection method using Transformer neural network is proposed. The proposed method combines wavelet packet transform time-frequency domain energy features and fast Fourier transform frequency domain features to generate input feature matrices, and solve the input problem of the Transformer neural network. At the same time, a normalization positional encoding method applied to rolling bearing fault type detection is proposed to solve the position coding problems of Transformer neural network in the field of rolling bearing fault analysis. A method for processing the input feature matrix from the bidirectional by the Transformer neural network and an error sample weight enhancement mechanism during the model training process are proposed to improve the robustness of the proposed method. The detection performance of the proposed method is verified using the rolling bearing dataset of the KAt data center. Compared with existing deep learning methods, the proposed method has certain advantages in both time efficiency and accuracy performance. The accuracy can be improved by more than 11%, and the average processing time of each sample is less than 1ms. |
|
查看全文
查看/发表评论 下载PDF阅读器 |
| 关闭 |
|
|
|
