| 摘 要: 减速箱对拉丝辊的转速固定有重要作用,由于拉丝设备结构紧密,内部零件的运行状态不易于观察,因此减速箱轮齿故障导致的转速配比异常很难被及时发现,针对拉丝机减速箱存在的故障诊断环节缺失问题,提出一种遗传算法与优化注意力模块改进的残差网络的故障诊断方法。首先,通过小波包分解与带通滤波的混合方法清洗数据,依照生产车间实际情况提出综合评价指标,并按照指标需求选择小波包分解层数;其次,针对残差网络与注意力模块进行改进;最后,将经过连通域分析与二值化后的特征图送入改进后的模型进行诊断。结果表明,该方法的诊断准确率比注意力-残差网络模型(Squeeze-and-Excitation-ResNet, SE-ResNet)提升了7.32%,比卷积神经网络-极限学习机模型(Convolutional Neural Network-Extreme Learning Machine, CNN-ELM)提升了8.81%,针对注意力模块(Squeeze-and-Excitation Module, SE)的改进将模型的单次诊断时间在原来的基础上缩短0.92 s,对塑编拉丝车间中减速箱的维护具有较大的实用价值。 |
| 关键词: 故障诊断;深度学习;遗传算法;挤压-激励模块;拉丝机;残差网络 |
|
中图分类号: TP206.3
文献标识码: A
|
| 基金项目: 浙江省科技计划项目(2022C01065);浙江省基础公益研究计划项目(LGG21E050024) |
|
| Fault Diagnosis of Wire Drawing Machine Reducer Based on Improved Residual Network |
|
ZOU Zhicheng1, WAN Changjiang1,2, RU Xin1
|
(1.School of Mechanical Engineering, Zhejiang Sci-Tech University, Hangzhou 310018, China;
2.Research Institute of Zhejiang Sci-Tech University in Longgang, Wenzhou 325000, China)
Mixier1130@163.com; wanchj@zstu.edu.cn; ruxin@zstu.edu.cn
|
| Abstract: The reducer plays an important role in fixing the speed of the wire drawing roller. Due to the compact structure of the drawing equipment and the difficulty in observing the operating status of internal parts, abnormal speed ratio caused by gearbox tooth faults is difficult to be found in time. To address the problem of missing fault diagnosis links in the reducer of the wire drawing machine, this paper proposes a fault diagnosis method based on genetic algorithm and improved residual network optimization attention module. Firstly, the data are cleaned by a hybrid method of wavelet packet decomposition and band-pass filtering. Comprehensive evaluation indicators are proposed based on the actual situation of the production workshop, and the number of wavelet packet decomposition layers is selected according to the indicator requirements. Secondly, the residual network and attention module are improved. Finally, the feature map after connected domain analysis and binarization is sent to the improved model for diagnosis. The results show that the diagnostic accuracy of the proposed method is 7.32% higher than that of traditional SEResNet (SE: Squeeze-and-Excitation), and 8. 81% higher than that of CNN-ELM (Convolutional Neural Network-Extreme Learning Machine). The improvement of SE module shortens the single diagnosis time of the model by 0.92 s, which has great practical value for the maintenance of the reducer in the plastic weaving and drawing workshop. |
| Keywords: fault diagnosis; deep learning; genetic algorithm; Squeeze-and-Excitation module; wire drawing machine; residual network |
