Page 108 - 《摩擦学学报》2021年第1期

第 41 卷 第 1 期 摩 擦 学 学 报 Vol 41 No 1
2021 年 1 月 Tribology Jan, 2021

DOI: 10.16078/j.tribology.2020020

基于Mask R-CNN网络的磨损颗粒

智能识别与应用

*
杨智宏, 贺石中 , 冯 伟, 李秋秋, 何伟楚
(广州机械科学研究院有限公司，广东 广州 510000)

摘 要: 针对设备磨损故障诊断中磨粒识别技术难度高、工作主观经验影响大等问题，采用深度学习技术开展了磨
粒智能识别的研究，提出了基于Mask R-CNN卷积神经网络的磨粒数字化表征方法. 该方法利用迁移学习训练基于
Mask R-CNN网络的磨粒识别模型对图像中磨粒进行识别和实例分割，然后使用Suzuki85算法、迭代算法、等比例计
算方法计算出磨粒的真实尺寸，解决了磨粒分析中难定量分析的问题. 结果表明：基于Mask R-CNN网络(采用R-
101-FPN骨干网络)训练的磨粒识别模型可以对图像中多个异常磨损颗粒进行识别，综合准确率和召回率达到当前
图像识别领域的主流水平. 辅以上述Suzuki85等算法，成功实现磨粒图像的定量评价分析，对促进设备故障诊断技
术的自动化发展和工业应用具有一定的实际应用价值.
关键词: 卷积神经网络; 深度学习; Mask R-CNN; 磨粒识别; 磨粒分析
中图分类号: TH117.1 文献标志码: A 文章编号: 1004-0595(2021)01–0105–10

Intelligent Identification of Wear Particles Based on Mask R-

CNN Network and Application

*
YANG Zhihong, HE Shizhong , FENG Wei, LI Qiuqiu, HE Weichu
(Guangzhou Mechanical Engineering Research Institute, Co Ltd, Guangdong Guangzhou 510000, China)
Abstract: In this paper, we presented a digital characterization method of abrasive particles based on deep learning and
Mask R-CNN convolutional neural network that enabled us to solve the problem in equipment wear fault diagnosis such
as high difficulty of abrasive particle identification and great influence of subjective experience. This method was used
to transfer learning of training the wear particle recognition model based on the Mask R-CNN network to identify and
segment the wear particles in the image, and then using the Suzuki85 algorithm, iterative algorithm, and proportional
calculation to calculate the true size of the wear particles. It solved the problem of difficult quantitative analysis in
abrasive particle analysis. The experimental results showed that the wear particle recognition model based on the Mask
R-CNN network (using the R-101-FPN backbone network) can identify multiple abnormal wear particles in the image,
and the comprehensive accuracy rate and recall rate came up to mainstream standard level of image recognition.
Supplemented by the above algorithm, it successfully implemented quantitative evaluation and analysis of wear images,
and was practical and valuable for promoting the automatic development and industrial application of equipment wear
fault diagnosis.
Key words: convolutional neural network; deep learning; mask R-CNN; wear particle recognition; wear particle analysis

Received 5 February 2020, revised 10 June 2020, accepted 23 June 2020, available online 28 January 2021.
*Corresponding author. E-mail: prof__heshizhong@163.com, Tel: +86-20-32389529.
The project was supported by the National Key Research and Development Project (2018YFB2001604), International Cooperation
Project of Guangzhou Development Zone (2018GH12) and Post-doctoral Program of Guangzhou Mechanical Engineering Research
Institute, Co, Ltd (17300065).
国家重点研发计划(2018YFB2001604)，广州市开发区国际合作项目(2018GH12)和广州机械科学研究院有限公司博士后专项
(17300065)资质.