基于异构集成学习的棉纱纱疵定量分析方法

doi:10.13475/j.fzxb.20220100501

摘要/Abstract

摘要：

为解决当前棉纱纱疵定量分析方法精度低、可靠性差的问题,提出一种基于异构集成学习的纱疵定量分析方法。首先,建立基于电容传感器的纱疵检测二维动态仿真模型,分析纱疵尺寸对纱疵信号的影响规律。其次,针对非平稳、非线性的纱疵信号难处理以及纱疵量化特征不明显的问题,采用时域分析方法,提取纱疵信号的时域参数作为纱疵定量分析特征,针对传统阈值和数值分析方法对纱疵定量精度低的问题,以支持向量机回归算法和径向基神经网络算法组成初级学习器,集成梯度提升决策树为元学习器,建立用于纱疵定量分析的异构集成学习算法模型。实验结果表明,本文方法比其它单模型回归拟合方法的检测准确率提升约10%,验证了本文方法对纱疵定量分析结果的可靠性。

关键词: 棉纱, 纱疵, 纱疵定量分析, 纱疵信号, 纱线质量, 异构集成学习算法

Abstract:

Objective Yarn defects are an important index to evaluate yarn quality, and it is necessary to classify the common yarn defects such as coarse yarn segment, thin yarn segment and neps in a more detailed way to achieve multilevel management of yarn quality. The key to yarn defect grading is how to quantitatively analyze the appearance geometric dimensions such as yarn defect length and diameter. Aiming at the problems of low accuracy and poor reliability of the current quantitative analysis methods for cotton yarn defects, a quantitative analysis method for cotton yarn defects based on heterogeneous integrated learning is proposed.

Methods A two-dimensional dynamic simulation model for yarn defect detection based on capacitance sensor was established to analyze the influence law of yarn defect size on its signal. Following that, the time domain analysis method was adopted to extract the time domain parameters of yarn defect signals as the quantitative analysis characteristics for yarn defects. Then, the support vector regression algorithm, radial basis function neural network algorithm and gradient boosting decision tree were used as primary and meta learners 'respectively' to establish a heterogeneous integrated learning algorithm model for quantitative analysis of yarn defects, which was verified by experiments finally.

Results Based on the capacitive sensor, the dynamic simulation modeling and analysis of yarn defect detection was carried out, and the comprehensive capacitive sensor detection circuit principle and dynamic simulation analysis results showed that the yarn defect length and yarn defect diameter were the key factors affecting the yarn defect signal change, among which the peak change of the yarn defect signal was affected by the superposition coupling of the yarn defect diameter and the yarn defect length, and the duration of the yarn defect signal peak was only affected by the yarn defect length. The results collected by the time domain analysis method showed that the time domain feature parameters of different levels of yarn defects deminstrated obvious differences, proving that the time domain feature parameters could be adopted to estimate the appearance geometry of yarn defects. However, the relationship between the time domain feature parameters and the appearance geometry of the yarn blemishes remained vague and nonlinear. Therefore, a network with strong nonlinear approximation capability was required to map the time domain parameters and the appearance geometry of yarn defects, namely the quantitative analysis algorithm of yarn defects based on heterogeneous integration learning. The root mean square error and mean absolute error of the yarn defect diameter test set were 0.002 1 and 0.001 2, and the root mean square error and mean absolute error of the yarn defect length test set were 0.002 6 and 0.001 4, which represents a greater improvement than other types of single-model algorithms, and the goodness of fit was close to 1.00, which fully demonstrates that the algorithm proposed in this paper has a better fitting effect on the yarn defect diameter and yarn defect length, and the model has stronger reliability.

Conclusion A quantitative analysis method of yarn defects based on heterogeneous ensemble learning is proposed. The method picked the yarn defects signal by capacitive sensor, combined the time domain characteristic parameter extraction algorithm and multi-model heterogeneous ensemble algorithm, and conducted quantitative analysis of non-stationary and nonlinear yarn defects signal. The experimental results confirmed that the yarn defect quantitative analysis model based on integration of heterogeneous learning can improve the appearance of yarn defects quantitative accuracy of geometry size with the method of optimal fitting R² close to 1.00. Compared with the conventional single model algorithm, the accuracy is improved by 10%, indicating the new method has good generalization ability and stability. It provides a more effective quantitative analysis scheme for yarn defects detection system based on electrical signal.

Key words: cotton yarn, yarn defect, quantitative analysis of yarn defect, yarn defect signal, yarn quality, heterogeneous ensemble learning algorithm

中图分类号:

TS111.9

杨芸, 孙通, 梁振宇, 彭广, 鲍劲松. 基于异构集成学习的棉纱纱疵定量分析方法[J]. 纺织学报, 2023, 44(05): 93-101.

YANG Yun, SUN Tong, LIANG Zhenyu, PENG Guang, BAO Jinsong. Quantitative analysis method of cotton yarn defects based on heterogeneous ensemble learning[J]. Journal of Textile Research, 2023, 44(05): 93-101.

图/表 15

图1

图2

图3

图4

图5

图6

表1

纱疵信号时域特征参数"

纱疵级别	长度/mm	直径/mm	X_max	X_rms	X_min	L_t	S²	$X ˉ$
A2	8	0.569	1.28	0.23	-0.26	22	0.05	0.06
A3	8	0.602	1.48	0.28	-0.22	25	0.08	0.10
B2	12	0.593	1.34	0.26	-0.28	35	0.09	0.12
C3	20	0.647	1.60	0.35	-0.32	50	0.12	0.00
F	120	0.296	0.64	0.18	-0.18	145	0.03	0.05
H1	115	0.136	0.08	0.09	-0.31	138	0.01	-0.06

表1

图7

图8

图9

图10

表2

纱疵外观几何尺寸定量分析时域特征参数"

组编号	纱疵类型	长度/mm	直径/mm	X_max	X_rms	X_min	L_t	S²	$X ˉ$
1	棉结	8	0.569	1.28	0.23	-0.26	22	0.05	0.06
2	棉结	8	0.602	1.48	0.28	-0.22	25	0.08	0.1
3	短粗节	12	0.593	1.34	0.26	-0.28	35	0.09	0.12
…		…	…	…	…	…	…	…	…
200	短粗节	20	0.647	1.60	0.35	-0.32	50	0.12	0.00
201	短粗节	46	0.403	0.88	0.26	-0.40	90	0.07	0.01
…		…	…	…	…	…	…	…	…
599	长粗节	120	0.296	0.64	0.18	-0.18	145	0.03	0.05
600	长细节	115	0.136	0.08	0.09	-0.31	138	0.01	-0.06

表2

图11

图12

表3

参考文献 17

[1]	王充, 赵帆. 纱线疵点产生原因分析[J]. 棉纺织技术, 2016, 44(10):56-58.
	WANG Chong, ZHAO Fan. Analysis on the cause of yarn defect[J]. Cotton Textile Technology, 2016, 44(10): 56-58.
[2]	玉红. 电子清纱器与成纱质量[J]. 黑龙江纺织, 2005(4):23-25.
	YU Hong. Electronic yarn cleaner and yarn quality[J]. Heilongjiang Textile, 2005 (4):23-25.
[3]	黄小强. 一种智能清纱算法的研究与实现[D]. 杭州: 杭州电子科技大学, 2011:17-23.
	HUANG Xiaoqiang. Research and implementation of an intelligent yarn clearing algorithm[D]. Hangzhou: Hangzhou Dianzi University, 2011: 17-23.
[4]	PINTO J G, CARVALHO V, MONTEIRO J L, et al. Yarn-mass measurement with 1-mm-length samples[J]. IEEE Transactions on Industrial Electronics, 2007, 54(2): 1177-1183. doi: 10.1109/TIE.2007.893051
[5]	盛国俊, 董永贵. 纱线质量光电信号的消噪处理和异常检测[J]. 清华大学学报(自然科学版), 2010, 50(2): 229-231.
	SHENG Guojun, DONG Yonggui. Yarn quality photoelectric signal denoising and abnormal detec-tion[J]. Journal of Tsinghua University (Science & Technology), 2010, 50 (2):229-231.
[6]	李变变, 李忠健, 潘如如, 等. 纱线条干均匀性序列图像测量方法[J]. 纺织学报, 2016, 37(11):26-31.
	LI Bianbian, LI Zhongjian, PAN Ruru, et al. Measurement of yarn evenness using sequence images[J]. Journal of Textile Research, 2016, 37 (11):26-31.
[7]	KANDROODI M R, ARAABI B N, BASSIRI M M, et al. Estimation of depth and length of defects from magnetic flux leakage measurements:verification with simulations, experiments, and pigging data[J]. IEEE Transactions on Magnetics, 2017, 53(3):1-10.
[8]	GE Yu, LIU Jinhai, ZHANG Huaguang, et al. An iterative stacking method for pipeline defect inversion with complex MFL signals[J]. IEEE Transactions on Instrumentation and Measurement, 2020, 69(6):3780-3788. doi: 10.1109/TIM.19
[9]	WANG Zhenwei, FEI Yuan, YE Pengxin, et al. Crack characterization in ferromagnetic steels by pulsed eddy current technique based on GA-BP neural network model[J]. Journal of Magnetism and Magnetic Materials, 2020:2-8.
[10]	WOLPERT D H. Stacked generalization[J]. Neural Networks, 1992, 5(2): 241-259. doi: 10.1016/S0893-6080(05)80023-1
[11]	ALVES A. Stacking machine learning classifiers to identify higgs bosons at the LHC[J]. Journal of Instrumentation, 2017, 12(5):1-3.
[12]	YANG Jiangji, LI Jiangqiang, SHEN Ruifang, et al. Exploiting ensemble learning for automatic cataract detection and grading[J]. Computer Methods and Programs in Biomedicine, 2016, 124: 45-57. doi: 10.1016/j.cmpb.2015.10.007 pmid: 26563686
[13]	王佶宣, 邓斌, 王江. 基于经验模态分解与RBF神经网络的短期风功率预测[J]. 电力系统及其自动化学报, 2020, 32(11):109-115.
	WANG Jixuan, DENG Bin, WANG Jiang. Short-term wind power prediction based on empirical mode decomposition and RBF neural network[J]. Power Systems and Automation, 2020, 32 (11):109-115.
[14]	周立俊, 黄腾, 王思捷, 等. 基于GA-SVR的地铁隧道沉降预测[J]. 地理空间信息, 2021, 19(3):115-117.
	ZHOU Lijun, HUANG Teng, WANG Sijie, et al. Subway tunnel settlement prediction based on GA-SVR[J]. Geospatial Information, 2021, 19(3):115-117.
[15]	KE G L, MENG Q, FINLEY T, et al. Lightgbm: a highly efficient gradient boosting decision tree[J]. Advances in Neural Information Processing Systems, 2017, 30(2): 3146-3154.
[16]	FRIEDMAN J. Greedy function approximation: a gradient boosting machine[J]. Annals of Statistics, 2001, 29(5):1189-1232. doi: 10.1214/aos/1013203450
[17]	孙通, 杨芸, 杨耀, 等. 基于时频特征学习的分步式纱疵检测方法[J]. 东华大学学报(自然科学版), 2022, 48(5):25-34.
	SUN Tong, YANG Yun, YANG Yao, et al. Step-by-step yarn defect detection method based on time-frequency feature learning[J]. Journal of Donghua University(Natural Science), 2022, 48(5):25-34.

Metrics

Viewed

Full text

Abstract

Cited

Shared

Discussed

模型	纱疵几何尺寸	R²	X_mae	X_rmse	时间/s
本文方法	D	0.993	0.001 2	0.002 1	0.199 4
本文方法	L_y	0.989	0.001 4	0.002 6	0.199 4
GBDT	D	0.962	0.001 2	0.058 6	0.249 7
GBDT	L_y	0.945	0.010 2	0.055 6	0.235 9
KNN	D	0.924	0.009 6	0.041 5	0.009 9
KNN	L_y	0.902	0.068 9	0.379 6	0.009 8
XGBoost	D	0.963	0.045 8	0.068 8	0.154 2
XGBoost	L_y	0.925	0.064 5	0.043 2	0.162 3
RBF	D	0.936	0.005 2	0.078 9	0.101 9
RBF	L_y	0.922	0.052 3	0.081 2	0.102 3
SVR	D	0.936	0.005 7	0.064 5	0.149 8
SVR	L_y	0.925	0.049 5	0.078 6	0.148 6