Soft sensor of rare earth element content with transfer learning and residual attention convolutional neural network

doi:10.11949/0438-1157.20241195

Abstract

Abstract:

Online detection of rare earth element content is a key link in rare earth industrial process control. Aiming at the problem that the existing soft measurement model of single color feature is not ideal, a soft measurement method of rare earth element content based on transfer learning residual attention convolutional network is proposed. Initially, prominent features such as color and texture are extracted from images of rare earth solutions. Additionally, latent convolutional features along with other critical elements are utilized as inputs to the soft sensing model. Subsequently, we design a one-dimensional CNN featuring multiple residual attention blocks to accommodate the one-dimensional nature of the rare earth solution image features. An attention mechanism is integrated, enabling the model to self-adjust the weighting of features based on their contribution, thereby enhancing model accuracy. The inclusion of a residual structure addresses issues related to vanishing or exploding gradients effectively. To make full use of solution image data in production process and reduce sample collection, a transfer learning strategy is employed. This strategy leverages data and knowledge accumulated from a source task, the maximum mean difference is used to measure the difference of feature distribution between the source domain and the target domain data, then the migration level and parameters are determined, and substantially improving the training outcomes of the target network. Finally, based on the laboratory image acquisition device and combined with field data, the simulation validation was conducted, and the results demonstrate the effectiveness of the proposed method.

Key words: rare earth element concentrations, one-dimensional convolutional neural network, transfer learning, residual attention mechanism

摘要：

稀土元素组分含量在线检测是稀土工业过程控制的关键环节。针对现有单一颜色特征的软测量模型效果不太理想的问题，提出一种融合迁移学习残差注意力卷积网络的稀土元素组分含量软测量方法。首先，提取稀土溶液颜色和纹理等显性特征，同时考虑稀土图像隐性卷积特征和其他关键特征作为该模型输入量；其次，针对稀土溶液图像特征具有一维特性，设计融合多残差注意力块的一维卷积神经网络，有效改善模型网络梯度消失或梯度爆炸问题，使模型能根据特征贡献度为特征自适应分配权重并提高模型预测精度；接着，为充分利用产线各级溶液图像数据，减少样本数量采集，加入迁移学习策略，在源任务中积累模型知识，通过最大均值差异衡量源域与目标域数据的特征分布差异，确定迁移层级及参数，显著增强目标网络的训练效果；最后，基于现场采集的稀土溶液图像和化验数据进行仿真对比实验，结果表明所提方法的有效性。

关键词: 稀土组分含量, 一维卷积神经网络, 迁移学习, 残差注意力机制

CLC Number:

TF 845

Fangping XU, Hui YANG, Jun CHEN, Jianyong ZHU, Rongxiu LU. Soft sensor of rare earth element content with transfer learning and residual attention convolutional neural network[J]. CIESC Journal, 2025, 76(4): 1647-1660.

徐芳萍, 杨辉, 陈俊, 朱建勇, 陆荣秀. 基于迁移学习与残差注意力卷积网络的稀土元素组分含量软测量[J]. 化工学报, 2025, 76(4): 1647-1660.

Figures/Tables 18

References 34

1	Chai T Y, Yang H. Situation and developing trend of rare-earth countercurrent extraction processes control[J].Journal of Rare Earths, 2004, 5(1): 604-610.
2	Zhang N, Huang C Z, Hu B. ICP-AES determination of trace rare earth elements in environmental and food samples by on-line separation and preconcentration with acetylacetone-modified silica gel using microcolumn[J]. Analytical Sciences, 2007, 23(8): 997-1002.
3	Tan D C, Zhu J M, Wang X L, et al. High-sensitivity determination of Cd isotopes in low-Cd geological samples by double spike MC-ICP-MS[J]. Journal of Analytical Atomic Spectrometry, 2020, 35(4): 713-727.
4	Zawisza B, Pytlakowska K, Feist B, et al. Determination of rare earth elements by spectroscopic techniques: a review[J]. Journal of Analytical Atomic Spectrometry, 2011, 26(12): 2373-2390.
5	袁静, 沈加林, 刘建坤, 等. 高能偏振能量色散X射线荧光光谱仪测定地质样品中稀土元素[J]. 光谱学与光谱分析, 2018, 38(2): 582.
	Yuan J, Shen J L, Liu J K, et al. Determination of rare earth elements in geological samples by high-energy polarized energy-dispersive X-ray fluorescence spectrometry[J]. Spectroscopy and Spectral Analysis, 2018, 38(2): 582.
6	杨辉, 高子洁, 陆荣秀. 基于稀土离子颜色特征识别的组分含量检测方法[J]. 中国稀土学报, 2012, 30(1): 108-112.
	Yang H, Gao Z J, Lu R X. Detection method of component content based on rare earth ions color characteristics identification[J]. Journal of the Chinese Society of Rare Earths, 2012, 30(1): 108-112.
7	Lu R X, Yang H. Soft measurement for component content based on adaptive model of Pr/Nd color features[J]. Chinese Journal of Chemical Engineering, 2015, 23(12): 1981-1986.
8	陆荣秀, 叶兆斌, 杨辉, 等. 镨/钕萃取过程组分含量多RBF模型预测[J]. 化工学报, 2016, 67(3): 974-981.
	Lu R X, Ye Z B, Yang H, et al. Multi-RBF models based prediction of component content for Pr/Nd extraction process[J]. CIESC Journal, 2016, 67(3): 974-981.
9	朱建勇, 张旭乾, 杨辉, 等. 单光照条件变化的镨/钕元素组分含量软测量[J]. 化工学报, 2019, 70(2): 780-788.
	Zhu J Y, Zhang X Q, Yang H, et al. Soft-sensing of Pr/Nd component content under different single illumination conditions[J]. CIESC Journal, 2019, 70(2): 780-788.
10	陆荣秀, 何权恒, 杨辉, 等. 基于GA-ELM的稀土混合溶液多组分含量预测[J]. 计算机工程, 2021, 47(1): 284-290, 297.
	Lu R X, He Q H, Yang H, et al. Multi-component content prediction of rare earth mixed solution based on GA-ELM[J]. Computer Engineering, 2021, 47(1): 284-290, 297.
11	陆荣秀, 邓彪, 杨辉, 等. 基于改进的GRA-即时学习算法的镨/钕元素组分含量预测[J]. 控制与决策, 2024, 39(2): 458-466.
	Lu R X, Deng B, Yang H, et al. Prediction of Pr/Nd component content based on improved GRA-just-in-time learning algorithm[J]. Control and Decision, 2024, 39(2): 458-466.
12	Shao L, He J W, Lu X, et al. Aircraft skin damage detection and assessment from UAV images using GLCM and cloud model[J]. IEEE Transactions on Intelligent Transportation Systems, 2024, 25(3): 3191-3200.
13	Yin S H, Shao Y J, Wu A X, et al. Texture features analysis on micro-structure of paste backfill based on image analysis technology[J]. Journal of Central South University, 2018, 25(10): 2360-2372.
14	Varish N, Hasan M K, Khan A, et al. Content-based remote sensing image retrieval method using adaptive tetrolet transform based GLCM features[J]. Journal of Intelligent & Fuzzy Systems, 2023, 44(6): 9627-9650.
15	Alibabaei S, Rahmani M, Tahmasbi M, et al. Evaluating the gray level co-occurrence matrix-based texture features of magnetic resonance images for glioblastoma multiform patients' treatment response assessment[J]. Journal of Medical Signals and Sensors, 2023, 13(4): 261-271.
16	Kamilaris A, Prenafeta-Boldú F X. Deep learning in agriculture: a survey[J]. Computers and Electronics in Agriculture, 2018, 147: 70-90.
17	王李管, 陈斯佳, 贾明滔, 等. 基于深度学习的黑钨矿图像识别选矿方法[J]. 中国有色金属学报, 2020, 30(5): 1192-1201.
	Wang L G, Chen S J, Jia M T, et al. Beneficiation method of wolframite image recognition based on deep learning[J]. The Chinese Journal of Nonferrous Metals, 2020, 30(5): 1192-1201.
18	Liu E Y, Zhao J X, Sofeia N. Students' entire deep learning personality model and perceived teachers' emotional support[J]. Frontiers in Psychology, 2022, 12: 793548.
19	Jiang R H. Understanding, Investigating, and promoting deep learning in language education: a survey on Chinese college students' deep learning in the online EFL teaching context[J]. Frontiers in Psychology, 2022, 13: 955565.
20	Sabab M N, Chowdhury M A R, Nirjhor S M M I, et al. Bangla speech recognition using 1D-CNN and LSTM with different dimension reduction techniques[M]//Emerging Technologies in Computing. Cham: Springer International Publishing, 2020: 158-169.
21	Ozcanli A K, Baysal M. Islanding detection in microgrid using deep learning based on 1D CNN and CNN-LSTM networks[J]. Sustainable Energy, Grids and Networks, 2022, 32: 100839.
22	Du R Y, Zhu S J, Ni H J, et al. Valence-arousal classification of emotion evoked by Chinese ancient-style music using 1D-CNN-BiLSTM model on EEG signals for college students[J]. Multimedia Tools and Applications, 2023, 82(10): 15439-15456.
23	Matsuura K, Ashihara T, Moriya T, et al. Transfer learning from pre-trained language models improves end-to-end speech summarization[C]//INTERSPEECH 2023. ISCA, 2023, 23(6): 04233.
24	Yang C H, Liang H P, Huang K K, et al. A robust transfer dictionary learning algorithm for industrial process monitoring[J]. Engineering, 2021, 7(9): 1262-1273.
25	Wadud M A H, Mridha M F, Shin J, et al. Deep-BERT: transfer learning for classifying multilingual offensive texts on social media[J]. Computer Systems Science and Engineering, 2023, 44(2): 1775-1791.
26	陆荣秀. 基于颜色特征的镨/钕萃取过程组分含量检测方法研究[D]. 南昌: 南昌大学, 2015.
	Lu R X. Study on detection method of component content in Praseodymium/Neodymium extraction process based on color characteristics[D]. Nanchang: Nanchang University, 2015.
27	He K M, Zhang X Y, Ren S Q, et al. Deep residual learning for image recognition[C]//2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 2016: 770-778.
28	刘铄, 宋俊材, 陆思良, 等. 基于灰度纹理特征提取和CS-SNN的双初级永磁同步直线电机退磁故障诊断研究[J]. 中国电机工程学报, 2023, 43(16): 6464-6474.
	Liu S, Song J C, Lu S L, et al. Demagnetization fault diagnosis research of DPPMSLM based on gray texture feature extraction and CS-SNN[J]. Proceedings of the CSEE, 2023, 43(16): 6464-6474
29	Özkan K, Mert A, Özdemir S. A new proposed GLCM texture feature: modified Rényi Deng entropy[J]. The Journal of Supercomputing, 2023, 79(18): 21507-21527.
30	Breiman L. Random forests[J]. Machine Learning, 2017, 45: 5-32.
31	Kamilaris A, Prenafeta-Boldú F X. A review of the use of convolutional neural networks in agriculture[J]. The Journal of Agricultural Science, 2018, 156(3): 312-322.
32	唐杰, 罗彦波, 李翔宇, 等. 基于近红外光谱数据的一维卷积神经网络模型研究[J]. 光谱学与光谱分析, 2024, 44(3): 731-736.
	Tang J, Luo Y B, Li X Y, et al. Study on one-dimensional convolutional neural network model based on near-infrared spectroscopy data[J]. Spectroscopy and Spectral Analysis, 2024, 44(3): 731-736.
33	王晋东. 迁移学习导论[M]. 北京: 电子工业出版社, 2021.
	Wang J D. Introduction to Transfer Learning[M]. Beijing: Publishing House of Electronics Industry, 2021.
34	张奇涵. 基于机器学习算法的稀土萃取过程元素组分含量在线预测研究[D]. 赣州: 江西理工大学, 2023.
	Zhang Q H. On-line prediction of element composition content in rare earth extraction process based on machine learning algorithm[D]. Ganzhou: Jiangxi University of Science and Technology, 2023.

序号	Layer	卷积核大小-通道数	输出大小
1	输入	—	18×1
2	特征融合层	—	12×1
3	注意力机制层	8×1-8	23× 8
4	RAB1	3×1-8	23× 8
5	RAB2	3×1-16	23×16
6	RAB3	3×1-32	23×32
7	RAB4	3×1-64	23×64
8	平均池化层	3×164	11×64
9	全连接层	—	256×1
10	输出	—	1×1

序号	Layer	卷积核大小-通道数	输出大小
1	输入	—	18×1
2	特征融合层	—	12×1
3	注意力机制层	8×1-8	23× 8
4	RAB1	3×1-8	23× 8
5	RAB2	3×1-16	23×16
6	RAB3	3×1-32	23×32
7	RAB4	3×1-64	23×64
8	平均池化层	3×164	11×64
9	全连接层	—	256×1
10	输出	—	1×1

模型	P(准确率)/%		AP（平均准确率）/%	MeanRE	MaxRE	RMSE	R²
模型	Pr	Nd	AP（平均准确率）/%	MeanRE	MaxRE	RMSE	R²
CNN	81.30	80.60	80.95	2.8367	18.5314	2.1562	0.9931
RAB-CNN	84.60	85.20	84.90	2.3687	16.8647	1.9935	0.9976
MRAB-CNN	88.40	89.20	88.80	0.9278	4.3278	0.8025	0.9998

模型	P(准确率)/%		AP（平均准确率）/%	MeanRE	MaxRE	RMSE	R²
模型	Pr	Nd	AP（平均准确率）/%	MeanRE	MaxRE	RMSE	R²
CNN	81.30	80.60	80.95	2.8367	18.5314	2.1562	0.9931
RAB-CNN	84.60	85.20	84.90	2.3687	16.8647	1.9935	0.9976
MRAB-CNN	88.40	89.20	88.80	0.9278	4.3278	0.8025	0.9998

迁移学习策略	MeanRE	MaxRE	RMSE	R²
策略1（无迁移学习）	0.8009	2.3040	0.7011	0.9992
策略2（参数全固定迁移）	1.8575	6.5490	1.9234	0.9946
策略3（参数全迁移）	0.5047	2.7161	0.4403	0.9997
策略4（残差块参数迁移）	0.6171	1.3897	0.5447	0.9995
策略5（卷积层参数迁移）	0.6973	4.6806	0.5139	0.9996
策略6（基于MMD迁移）	0.3618	1.2179	0.4204	0.9997