基于SimAM和SpinalNet的列车轮对踏面缺陷分类模型

doi:10.16265/j.cnki.issn1003-3033.2022.06.2563

中国安全科学学报 ›› 2022, Vol. 32 ›› Issue (6): 38-43.doi: 10.16265/j.cnki.issn1003-3033.2022.06.2563

基于SimAM和SpinalNet的列车轮对踏面缺陷分类模型

张昌凡¹(), 胡新亮²(), 何静¹, 刘建华²^,^**, 侯娜¹

1 湖南工业大学电气与信息工程学院,湖南株洲 412007
2 湖南工业大学轨道交通学院,湖南株洲 412007

收稿日期:2022-01-09 修回日期:2022-04-14 出版日期:2022-06-28 发布日期:2022-12-28
通讯作者: 刘建华
作者简介:
张昌凡 (1960—),男,湖北武汉人,博士,教授,主要从事非线性控制及应用等方面的研究。E-mail: zhangchangfan@263.net。
何静,教授
刘建华,副教授
基金资助:
国家自然科学基金资助(52172403); 国家自然科学基金资助(62173137); 湖南省自然科学基金资助(2021JJ30217); 湖南省教育厅资助项目(19A137)

Defect classification model for high-speed train wheelset treads based on SimAM and SpinalNet

ZHANG Changfan¹(), HU Xinliang²(), HE Jing¹, LIU Jianhua²^,^**, HOU Na¹

1 College of Electrical and Information Engineering, Hunan University of Technology, Zhuzhou Hunan 412007, China
2 College of Railway Transportation, Hunan University of Technology, Zhuzhou Hunan 412007, China

Received:2022-01-09 Revised:2022-04-14 Online:2022-06-28 Published:2022-12-28
Contact: LIU Jianhua

摘要/Abstract

摘要：

为解决小样本问题下轮对踏面缺陷分类难题,提出一种基于简单无参注意力模块(SimAM)和脊柱神经网络(SpinalNet)踏面缺陷分类模型。首先,预训练网络提取原始图像各个类别特征图;其次,在有限的训练样本下,利用SimAM提取对缺陷图像表示性更强的类别特征;然后,利用SpinalNet关联特征图的局部和整体语义,得到缺陷类别特征的强区分性表示;最后,以强区分性表示特征输入带有L2正则化的softmax分类器,得到分类结果。试验结果表明:小样本任务评估指标准确率1和准确率2分别为68.35%和100%,优于目前主流深度学习模型,能够有效分类轮对踏面缺陷从而避免列车安全事故发生。

关键词: 轮对踏面, 缺陷分类, 简单无参注意力模块(SimAM), 脊柱神经网络(SpinalNet), L2正则化

Abstract:

In order to address classification difficulty of tread defects for small sample tasks, a classification model based on SimAM and SpinalNet was proposed. Firstly, feature maps of each category of original images were extracted from pre-trained networks. Secondly, class features with stronger representation of defect images were extracted by using SimAM under limited training samples, local and overall semantics of feature maps were correlated by utilizing SpinalNet to obtain a strong distinguishing representation of defect class features. Finally, strong discriminating representation features were inputted to softmax classifier with L2 regularization, and classification results were obtained. The research shows that the accuracy rate of evaluation index of small sample tasks was 68.35% and 100%, respectively, which was better than current mainstream deep learning model.

Key words: wheelset tread, defect classification, simple parameter-free attention module(SimAM), spinal neural network (SpinalNet), l2 regularization

张昌凡, 胡新亮, 何静, 刘建华, 侯娜. 基于SimAM和SpinalNet的列车轮对踏面缺陷分类模型[J]. 中国安全科学学报, 2022, 32(6): 38-43.

ZHANG Changfan, HU Xinliang, HE Jing, LIU Jianhua, HOU Na. Defect classification model for high-speed train wheelset treads based on SimAM and SpinalNet[J]. China Safety Science Journal, 2022, 32(6): 38-43.

图/表 7

图1

表1

ResNet50结构

模块	模块1	模块2	模块3	模块4	模块5	输出
ResNet50	7×7,64, stride=2	$1 × 1, 64 3 × 3, 64 1 × 1, 256 × 3$	$1 × 1, 128 3 × 3, 128 1 × 1, 512 × 4$	$1 × 1, 256 3 × 3, 256 1 × 1, 1024 × 6$	$1 × 1, 512 3 × 3, 512 1 × 1, 2048 × 6$	GAP Fc Softmax
输出尺寸	112×112	56×56	28×28	14×14	7×7	1×1

表1

图2

表2

表3

图3

表4

参考文献 24

[1]	曹源, 温佳坤, 马连川. 重大疫情下的列车动态编组与调度[J]. 交通运输工程学报, 2020, 20(3):120-128.
	CAO Yuan, WEN Jiakun, MA Lianchuan. Dynamic marshalling and scheduling of trains in major epidemics[J]. Journal of Traffic and Transportation Engineering, 2020, 20(3): 120-128.
[2]	任禹谋, 张琦, 袁志明, 等. 安全约束条件下高铁站到发线运用优化研究[J]. 中国安全科学学报, 2021, 31(1): 179-185. doi: 10.16265/j.cnki.issn 1003-3033.2021.01.026
	REN Yumou, ZHANG Qi, YUAN Zhiming, et al. Research on optimization for utilization of arrival and departure tracks at high-speed railway station under condition of safety constraints[J]. China Safety Science Journal, 2021, 31(1): 179-185. doi: 10.16265/j.cnki.issn 1003-3033.2021.01.026
[3]	NENOV N, DIMITROV E, VASILEV V, et al. Sensor system of detecting defects in wheels of railway vehicles running at operational speed[C]. Proceedings of the 2011 34^th International Spring Seminar on Electronics Technology (ISSE), 2011: 577-582.
[4]	YUE Jianhai, QIU Zhengding, CHEN Boshi. Application of Wavelet transform to defect detection of wheelflats of railway wheels[C]. 6^th International Conference on Signal Processing, 2002:29-32.
[5]	BO Liang, IWNICKI S, BALL A, YOUNG A E, et al. Adaptive noise cancelling and time-frequency techniques for rail surface defect detection[J]. Mechanical Systems and Signal Processing, 2015, 54: 41-51.
[6]	YANG Lingxiao, ZHANG Ruyuan, LI Lida, et al. Simam: a simple, parameter-free attention module for convolutional neural networks[C]. International Conference on Machine Learning. PMLR, 2021: 11 863-11 874.
[7]	刘颖, 雷研博, 范九伦, 等. 基于小样本学习的图像分类技术综述[J]. 自动化学报, 2021, 47(2): 297-315.
	LIU Ying, LEI Yanbo, FAN Jiulun, et al. Survey on image classification technology based on small sample learning[J]. Acta Automatica Sinica, 2021, 47(2): 297-315.
[8]	HU Jie, SHEN Li,SUN Gang. Squeeze-and-excitation networks[C]. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2018: 7132-7141.
[9]	WANG Qilong, WU Bangyu, ZHU Pengfei, et al. ECA-net: efficient channel attention for deep convolutional neural networks[C]. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020: 11 531-11 539.
[10]	HOU Qibin, ZHOU Daquan, FENG Jiashi. Coordinate attention for efficient mobile network design[C]. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021: 13 713-13 722.
[11]	ZHANG Hu, ZU Keke, LU Jian, et al. EPSANet: an efficient pyramid split attention block on convolutional neural network[J]. Computer Vision and Pattern Recognition, 2021: Doi: 10.48550/arxiv.2105.14447. doi: 10.48550/arxiv.2105.14447
[12]	LIU Huajun, LIU Fuqiang, FAN Xinyi, et al. Polarized self-attention: towards high-quality pixel-wise regression[J]. Computer Vision and Pattern Recognition, 2021: Doi: 10.48550/arXiv.2107.00782. doi: 10.48550/arXiv.2107.00782
[13]	QIN Zequn, ZHANG Pengyi, WU Fei, et al. Fcanet: frequency channel attention networks[C]. Proceedings of the IEEE/CVF International Conference on Computer Vision, 2021: 783-792.
[14]	SAINI R, JHA N K, DAS B, et al. Ulsam: ultra-lightweight subspace attention module for compact convolutional neural networks[C]. Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, 2020: 1627-1636.
[15]	LIU Yichao, SHAO Zongru, TENG Yueyang, et al. NAM: normalization-based attention module[J]. Computer Vision and Pattern Recognition, 2021: Doi: 10.48550/arXiv.2111.12419. doi: 10.48550/arXiv.2111.12419
[16]	KABIR H M, ABDAR M, JALALI S M J, et al. Spinalnet: deep neural network with gradual input[J]. Computer Vision and Pattern Recognition, 2022: Doi: 10.48550/arXiv.2007.03347. doi: 10.48550/arXiv.2007.03347
[17]	QASSIM H, VERMA A, FEINZIMER D. Compressed residual-VGG16 CNN model for big data places image recognition[C]. 2018 IEEE 8^th Annual Computing and Communication Workshop and Conference (CCWC), 2018: 169-175.
[18]	SHENG Tao, FENG Chen, ZHUO Shaojie, et al. A quantization-friendly separable convolution for mobilenets[C]. 2018 1^st Workshop on Energy Efficient Machine Learning and Cognitive Computing for Embedded Applications (EMC2), 2018: 14-18.
[19]	SANDLER M, HOWARD A, ZHU Menglong, et al. Mobilenetv2: inverted residuals and linear bottlenecks[C]. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2018: 4510-4520.
[20]	HE Kaiming, ZHANG Xiangyu, REN Shaoqing, et al. Deep residual learning for image recognition[C]. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2016: 770-778.
[21]	XIA Xiaoling, XU Cui, NAN Bing. Inception-v3 for flower classification[C]. 2017 2^nd International Conference on Image, Vision and Computing (ICIVC), 2017: 783-787.
[22]	BALDASSARRE F, MORI' N D G, et al. Deep koalarization: image colorization using cnns and inception-resnet-v2[J]. Computer Vision and Pattern Recognition, 2021: Doi: 10.48550/arXiv.1712.03400. doi: 10.48550/arXiv.1712.03400
[23]	CHOLLET F. Xception: deep learning with depthwise separable convolutions[C]. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2017: 1251-1258.
[24]	DOSOVITSKIY A, BEYER L, KOLESNIKOV A, et al. An image is worth 16x16 words: transformers for image recognition at scale[J]. Computer Vision and Pattern Recognition, 2020: Doi: 10.48550/arX-iv.2010.11929. doi: 10.48550/arX-iv.2010.11929

骨干网络	注意力机制	准确率1/%	准确率2/%	参数/ MByte
Resnet50	SE^[8]	60.76	98.73	25.70
	ECA^[9]	60.76	96.20	23.60
	CA^[10]	58.23	97.47	25.18
	EPSA^[11]	20.25	72.15	51.45
	PSA,模式P^[12]	59.49	98.73	31.99
	PSA,模式S^[12]	63.29	97.47	30.68
	FCA^[13]	55.70	94.94	24.65
	ULSAM^[14]	51.90	94.94	23.62
	NAM^[15]	58.23	92.41	23.61
	SimAM^[5]	64.56	100.00	23.60

骨干网络	预训练	SimAM	脊柱全连接	L2正则化	准确率1/%	准确率2/%	参数/MByte
ResNet50	—	—	—	—	41.77	94.94	23.60
	√	—	—	—	54.43	96.20	23.60
	√	√	—	—	64.56	100.00	23.60
	√	√	√	—	65.82	100.00	23.86
	√	√	√	√	68.35	100.00	23.86

模型	准确率 1/%	准确率2/%	参数/ MByte
VGG16^[17]	25.32	77.22	201.40
Mobilenetv1-&=1^[18]	18.99	89.87	3.24
Mobilenetv2^[19]	25.32	92.41	2.27
ResNet50^[20]	54.43	96.20	23.60
Inceptionv3^[21]	49.37	94.94	21.82
InceptionResNetV2^[22]	34.28	79.75	54.35
Xception^[23]	32.91	89.87	20.88
Vision Transformer^[24]	25.32	60.76	85.80
文中模型	68.35	100.00	23.86

基于SimAM和SpinalNet的列车轮对踏面缺陷分类模型

Defect classification model for high-speed train wheelset treads based on SimAM and SpinalNet

RichHTML

PDF

可视化

摘要/Abstract

引用本文

使用本文

图/表 7

参考文献 24

相关文章 1

编辑推荐

Metrics

本文评价