基于数据分块策略的双流多采样率GRU软测量建模

doi:10.11949/0438-1157.20251190

摘要/Abstract

摘要：

受传感器采样能力及工艺特性制约，一些生产过程变量通常具有不同的采样速率。然而，传统软测量建模一般假设所有数据采样速率相同，不能直接应用于多采样率工业场景。因此，提出一种基于数据分块策略的双流多采样率门控循环单元（Gated Recurrent Unit，GRU）的软测量建模方法。首先，基于采样频率高低设计数据分块策略，将数据划分为高频数据块与低频数据块；其次，构建双流特征提取网络，在低频数据流引入时序动态衰减模块，通过改进的衰减门控循环单元直接处理缺失数据，在高频数据流设计时序注意力增强模块，捕捉密集采样数据特征；进一步设计融合层自适应分配双流网络的隐藏特征权重，以构建处理多采样率数据的软测量模型。最后，通过脱丁烷塔过程及硫回收过程的应用仿真验证所提算法的预测效果。

关键词: 多采样率, 数据分块策略, 衰减门控循环单元, 软测量, 神经网络, 过程控制, 动态建模

Abstract:

Owing to limitations in sensor sampling capabilities and process characteristics, industrial process variables are often sampled at different rates. Conventional soft sensor modeling approaches, which typically assume uniform sampling rates across all variables, are not directly applicable to such multirate scenarios. To address this issue, this paper proposes a dual-stream modeling framework based on a data-block strategy and Gated Recurrent Units (GRUs). First, variables are segmented into high-frequency and low-frequency data blocks according to their sampling rates. A dual-stream network is then constructed for feature extraction: the low-frequency stream incorporates a temporal dynamic decay module, which employs a modified GRU with a decay mechanism to handle missing data points directly. In parallel, the high-frequency stream is equipped with a temporal attention enhancement module to capture discriminative features from the densely sampled data. Furthermore, a fusion layer is designed to adaptively aggregate the hidden features from both streams through learned weights, thereby building a robust soft sensor model capable of handling multirate data. The predictive performance and effectiveness of the proposed method are demonstrated through case studies on a debutanizer column process and a sulfur recovery unit process.

Key words: multirate sampling, data chunk strategy, decay gated recurrent unit, soft sensor, neural networks, process control, dynamic modeling

中图分类号:

TP273

张震, 史旭东, 王法正, 熊伟丽. 基于数据分块策略的双流多采样率GRU软测量建模[J]. 化工学报, DOI: 10.11949/0438-1157.20251190.

Zhen ZHANG, Xudong SHI, Fazheng WANG, Weili XIONG. A dual-stream multirate GRU soft sensor modeling based on data chunk strategy[J]. CIESC Journal, DOI: 10.11949/0438-1157.20251190.

图/表 19

图1 GRU结构图

Fig.1 GRU structure diagram

图2 DC-DSMGRU软测量模型结构图

Fig.2 Structure diagram of DC-DSMGRU soft measurement model

图3 数据分块策略说明图

Fig.3 Illustrative diagram of the Data-Chunk Strategy

图4 TDDGRU结构图

Fig.4 Structure diagram of TDDGRU

图5 TAEGRU结构图

Fig.5 Structure diagram of TAEGRU

表1 脱丁烷塔变量描述

Table 1 Variable description of debutanizer

变量名称	变量单位	采样间隔/min
塔顶温度	℃	10
塔顶压力	kg/cm²	10
回流流量	m³/h	20
出料流量	m³/h	20
第六层塔板温度	℃	10
塔底温度1	℃	10
塔底温度2	℃	10
丁烷浓度	mol%	40

表2 各多采样率模型丁烷浓度预测结果

Table 2 Prediction results of butane concentration in each multi-sampling rate model

所用模型	RMSE	MAE	R²
MLP	0.06341	0.05267	0.86452
RNN	0.05743	0.04662	0.90419
VPTN	0.05522	0.04573	0.92699
MRS-KNN	0.04831	0.03900	0.94412
T-LSTM	0.02848	0.02223	0.96792
DC-DSMGRU	0.02368	0.01793	0.98658

图6 多采样率模型对丁烷浓度预测曲线图

Fig.6 Prediction curve of butane concentration by multi-sampling rate model

表3 消融实验中丁烷浓度预测结果

Tab.3 Prediction results of butane concentration in ablation experiments

所用模型	RMSE	MAE	R²
GRU	0.05062	0.04389	0.93865
GRU-D	0.04873	0.04166	0.94315
TAEGRU	0.04307	0.03480	0.95559
TDDGRU	0.04351	0.03517	0.95468
DC-DSMGRU	0.02368	0.01793	0.98658

图7 消融实验中丁烷浓度预测误差频率分布图

Fig.7 Frequency distribution of butane concentration prediction error in ablation experiment

图8 消融实验预测误差箱线图

Fig.8 Box plot of prediction error of ablation experiment

图9 各超参数下丁烷预测浓度RMSE

Fig.9 Predicted concentration of RMSE of butane under each hyperparameter

图10 硫回收过程流程图

Fig.10 Flow chart of sulfur recovery process

表4 硫回收变量描述

Tab.4 Variable description of SRU

变量名称	变量单位	采样间隔/min
GAS气流	m³/h	10
AIR空气流	m³/h	10
AIR二次空气流	m³/h	20
SWS区域GAS气流	m³/h	20
SWS区域AIR空气流	m³/h	10
SO₂浓度	mg/m³	40

表5 各多采样率模型SO2浓度预测结果

Table 5 Prediction results of SO2 concentration in each multi-sampling rate model

所用模型	RMSE	MAE	R²
MLP	0.03242	0.02556	0.54825
VPTN	0.02843	0.02044	0.65253
RNN	0.02366	0.01870	0.75932
FCW-RNN	0.02136	0.01665	0.80388
T-LSTM	0.02034	0.01615	0.82210
DC-DSMGRU	0.01808	0.01393	0.85945

图11 SO2浓度预测曲线图

Fig.11 Prediction curve of SO2 concentration

图12 SO2浓度预测误差曲线图

Fig.12 Error curve of SO2 concentration prediction

表6 消融实验中SO2浓度预测结果

Tab.6 Prediction results of SO2 concentration in ablation experiment

所用模型	RMSE	MAE	R²
GRU	0.02503	0.01925	0.73081
GRU-D	0.02333	0.01792	0.76607
TAEGRU	0.02277	0.01707	0.77721
TDDGRU	0.02118	0.01562	0.80713
DC-DSMGRU	0.01812	0.01411	0.85884

图13 各超参数下SO2预测浓度RMSE

Fig.13 Predicted concentration of RMSE of SO2

参考文献 32

[1]	廖开继, 隋璘, 熊伟丽. 基于双阶段注意力的双流记忆调节GRU软测量建模[J]. 控制与决策, 2025, 40(9): 2848-2858.
	Liao K J, Sui L, Xiong W L. Soft sensor modeling based on dual-stream memory-modulated GRU with dual-stage attention mechanism[J]. Control and Decision, 2025, 40(9): 2848-2858.
[2]	李文华, 叶洪涛, 罗文广, 等. 基于MHSA-LSTM的软测量建模及其在化工过程中的应用[J]. 化工学报, 2024, 75(12): 4654-4665.
	Li W H, Ye H T, Luo W G, et al. Soft sensor modeling based on MHSA-LSTM and its application in chemical process[J]. CIESC Journal, 2024, 75(12): 4654-4665.
[3]	Zheng K Y, Funatsu K. Partial constrained least squares (PCLS) and application in soft sensor[J]. Chemometrics and Intelligent Laboratory Systems, 2018, 177: 64-73.
[4]	Yin S L, Li Y G, Sun B, et al. Mixed kernel principal component weighted regression based on just-in-time learning for soft sensor modeling[J]. Measurement Science and Technology, 2022, 33(1): 015102.
[5]	Zhu X L, Rehman K U, Wang B, et al. Modern soft-sensing modeling methods for fermentation processes[J]. Sensors, 2020, 20(6): 1771.
[6]	范龙飞, 史旭东, 熊伟丽. 跨工况下基于参数迁移的域适应宽度学习软测量建模[J]. 化工学报, 2025,76(9): 4644-4657.
	Fan L F, Shi X D, Xiong W L. Domain adaptive broad learning system with parameter transferring for cross-condition soft sensor modeling[J]. CIESC Journal, 2025, 76(9): 4644-4657.
[7]	Sun Q Q, Ge Z Q. A survey on deep learning for data-driven soft sensors[J]. IEEE Transactions on Industrial Informatics, 2021, 17(9): 5853-5866.
[8]	Purwono P, Ma'arif A, Rahmaniar W, et al. Understanding of convolutional neural network (CNN): a review[J]. International Journal of Robotics and Control Systems, 2022, 2(4): 739-748.
[9]	Xie R M, Jan N M, Hao K R, et al. Supervised variational autoencoders for soft sensor modeling with missing data[J]. IEEE Transactions on Industrial Informatics, 2020, 16(4): 2820-2828.
[10]	Pratama I, Permanasari A E, Ardiyanto I, et al. A review of missing values handling methods on time-series data[C]//2016 International Conference on Information Technology Systems and Innovation (ICITSI). October 24-27, 2016, Bandung, Indonesia. IEEE, 2017: 1-6.
[11]	Zhang S M, Zhao C H. Slow-feature-analysis-based batch process monitoring with comprehensive interpretation of operation condition deviation and dynamic anomaly[J]. IEEE Transactions on Industrial Electronics, 2019, 66(5): 3773-3783.
[12]	孙凯, 隋璘, 张芳芳, 等. 基于非负绞杀与长短期记忆神经网络的动态软测量算法[J]. 控制理论与应用, 2023, 40(1): 83-93.
	Sun K, Sui L, Zhang F F, et al. Dynamic soft sensor algorithm based on nonnegative garrote and long short-term memory neural network[J]. Control Theory & Applications, 2023, 40(1): 83-93.
[13]	Baytas I M, Xiao C, Zhang X, et al. Patient subtyping via time-aware LSTM networks[C]//Proceedings of the 23rd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining. Halifax NS Canada. ACM, 2017: 65-74.
[14]	Dey R, Salem F M. Gate-variants of gated recurrent unit (GRU) neural networks[C]//2017 IEEE 60th International Midwest Symposium on Circuits and Systems (MWSCAS). August 6-9, 2017, Boston, MA, USA. IEEE, 2017: 1597-1600.
[15]	王法正, 隋璘, 熊伟丽. 面向多采样率数据的TTPA-LSTM软测量建模[J]. 化工学报, 2025, 76(4): 1635-1646.
	Wang F Z, Sui L, Xiong W L. TTPA-LSTM soft sensor modeling for multi-sampling rate data[J]. CIESC Journal, 2025, 76(4): 1635-1646.
[16]	Che Z P, Purushotham S, Cho K, et al. Recurrent neural networks for multivariate time series with missing values[J]. Scientific Reports, 2018, 8: 6085.
[17]	Chang S C, Chen X, Zhao C H. Flexible Clockwork Recurrent Neural Network for multirate industrial soft sensor[J]. Journal of Process Control, 2022, 119: 86-100.
[18]	Chai Z, Zhao C H, Huang B. Variational progressive-transfer network for soft sensing of multirate industrial processes[J]. IEEE Transactions on Cybernetics, 2022, 52(12): 12882-12892.
[19]	Feng J, Li K Q. MRS-kNN fault detection method for multirate sampling process based variable grouping threshold[J]. Journal of Process Control, 2020, 85: 149-158.
[20]	钱子豪, 王磊, 靳晟. 基于Transformer-LSTM的高炉炼铁硅含量软测量方法[J]. 机电工程技术, 2025, 54(19): 35-40, 114.
	Qian Z H, Wang L, Jin S. Soft measurement method of silicon content in blast furnace ironmaking based on transformer-LSTM[J]. Mechanical & Electrical Engineering Technology, 2025, 54(19): 35-40, 114.
[21]	Ma L, Wang M W, Peng K X. A spatiotemporal industrial soft sensor modeling scheme for quality prediction with missing data[J]. IEEE Transactions on Instrumentation and Measurement, 2024, 73: 1-10.
[22]	冯路. 基于堆栈自编码器的加氢裂化多采样率数据软测量建模[D]. 长沙: 中南大学, 2022.
	Feng L. Soft sensor modeling for multi-rate data based on stacked autoencoder in the hydrocracking process[D]. Changsha: Central South University, 2022.
[23]	Ren L, Wang T, Laili Y, et al. A data-driven self-supervised LSTM-DeepFM model for industrial soft sensor[J]. IEEE Transactions on Industrial Informatics, 2022, 18(9): 5859-5869.
[24]	Tan Q X, Ye M, Yang B Y, et al. DATA-GRU: dual-attention time-aware gated recurrent unit for irregular multivariate time series[J]. Proceedings of the AAAI Conference on Artificial Intelligence, 2020, 34(1): 930-937.
[25]	熊中敏, 郭怀宇, 吴月欣. 缺失数据处理方法研究综述[J]. 计算机工程与应用, 2021, 57(14): 27-38.
	Xiong Z M, Guo H Y, Wu Y X. Review of missing data processing methods[J]. Computer Engineering and Applications, 2021, 57(14): 27-38.
[26]	Tzoumpas K, Estrada A, Miraglio P, et al. A data filling methodology for time series based on CNN and (Bi)LSTM neural networks[J]. IEEE Access, 2024, 12: 31443-31460.
[27]	Huang G L. Missing data filling method based on linear interpolation and lightgbm[J]. Journal of Physics: Conference Series, 2021, 1754(1): 012187.
[28]	Wu R, Hamshaw S D, Yang L, et al. Data imputation for multivariate time series sensor data with large gaps of missing data[J]. IEEE Sensors Journal, 2022, 22(11): 10671-10683.
[29]	王建军, 潘佳音, 赵珍珠, 等. 基于特征时间双注意力机制的短期光伏发电预测深度学习模型研究[J]. 智慧电力, 2025(4): 81-87.
	Wang J J, Pan J Y, Zhao Z Z, et al. A deep learning model for short-term photovoltaic power generation forecasting based on dual feature-temporal attention mechanism[J]. Smart Power, 2025(4): 81-87.
[30]	Lin B, Recke B, Knudsen J K H, et al. Data-driven soft sensor design with multiple-rate sampled data: a comparative study[C]//2007 European Control Conference (ECC). July 2-5, 2007, Kos, Greece. IEEE, 2015: 4740-4745.
[31]	潘建, 汪绪豪. 融合多尺度特征和注意力机制的时间序列预测模型[J/OL]. 计算机科学, 2025: 1-8. (2025-05-15).
	Pan J, Wang X H. Time series prediction model integrating multi-scale features and attention mechanism[J/OL]. Computer Science, 2025: 1-8. (2025-05-15).
[32]	Fortuna L, Rizzo A, Sinatra M, et al. Soft analyzers for a sulfur recovery unit[J]. Control Engineering Practice, 2003, 11(12): 1491-1500.