School of Artificial Intelligence and Computer Science, Jiangnan University, Wuxi 214000, China
Show less
文章历史+
Received
Published
Revised
2025-08-11
2026-08-18
2025-12-18
Online Date
2026-09-04
PDF (8387K)
摘要
周期性时序数据的高效分类在医疗、气象和工业生产等领域具有重要意义,然而复杂的现实环境(如传感器失灵)可能导致训练数据不足。为了开发一种高效的迁移学习方法,以提高随机向量函数链网络(random vector functional-link net,RVFL)在周期性时序数据不足情形下的分类性能,提出了基于输出迁移的随机向量函数链网络(output-based transfer learning for RVFL,OTL-RVFL)。其基本思想是利用随机向量函数链网络架构简洁、训练高效和性能出色的优势;首先以时间序列的周期为单位构建基础的RVFL分类模型,接着为其中数据不足的周期段建立OTL-RVFL模型;OTL-RVFL通过迁移前序周期的RVFL输出,并直接以解析解求得模型参数,增强对当前周期段时序数据的分类性能。依托公开数据集的实验结果表明,该方法将RVFL与迁移学习结合,有效提升了面对周期性时序数据不足时完成分类任务的性能,且具备显著的训练效率优势。
Abstract
Efficient classification of periodic time-series data has very important significance for various application fields like healthcare, meteorology and industrial production. However, the training data for such tasks may become insufficient due to complex circumstances like sensor failures. An output-based transfer learning for RVFL model (OTL-RVFL) was developed to improve random vector functional-link networks (RVFL) so as to achieve enhanced classification performance from insufficient periodic time-series data. The basic idea was to leverage RVFL as the base classification model, characterized by its simple architecture, efficient training and strong classification performance. Firstly, the basic RVFL classification model was built along each period for periodic time-series data. Secondly, for each period with insufficient data, an OTL-RVFL model was built to transfer the outputs of the corresponding basic RVFL models trained on one or more previous periods and learn the model parameters analytically in closed-form solution, thereby enhancing the classification performance on the current period’s time-series data. Experimental results on public datasets demonstrate that the proposed method effectively combines RVFL with transfer learning, significantly improving the classification performance for insufficient periodic time-series data classification while retaining RVFL’s remarkable advantages in training efficiency.
WANG Shi-tong, YANG Zi-lun.
Transfer Learning of Random Neural Networks for Time Series Data Classification[J].
Science Technology and Engineering, 2026, 26(23): 10050-10061 DOI:10.12404/j.issn.1671-1815.2505857
随着大数据与智能技术的迅猛发展,时间序列数据(time series data)的规模不断扩大,通过分析一段时序数据的特征并将它分配到已知类别中被称为时间序列分类(time series classification, TSC)[1]。时间序列数据(又称“时序数据”)广泛存在于医疗[2]、环保[3]和工业生产[4]等领域,其中相当一部分还呈现出明显的周期性特征。时间序列的周期性是指数据在特定时间间隔后会重复呈现相似的波动模式或规律性变化[5]。受制于复杂的现实环境(如传感器故障),周期性时序数据往往出现片段缺失,导致部分周期的有效数据不足。
随机向量函数链网络(random vector functional-link net,RVFL)[19-20]作为最经典的随机神经网络变体[21-22],凭借其简洁的架构、通用函数逼近能力[23-24]以及可解析解的模型训练方式,不仅拥有较好的时序数据分类性能,而且网络优化成本较低;直链的应用有效增强了RVFL的时序数据分析能力[25],得益于它的计算效率与性能优势,RVFL被广泛应用于分类[26]、回归[27]、聚类[28]以及高实时性要求[29-30]的诸多任务中[31]。毛琦等[32]提出了一种由非线性自回归结构与极限学习机组成的预测器和RVFL分类器构成的混合数据驱动诊断框架,引入故障特征叠加的诊断策略,在无需新增硬件的条件下显著提升了故障诊断的精度、效率和工程实用性。Malik等[33]提出基于线性回归的监督式扩展特征构建方法efRVFL(extended features RVFL),在输入空间中生成更具判别性的扩展特征,并集成多个在不同扩展特征空间上训练的efRVFL形成集成efRVFL框架,在未明显增加模型复杂度和训练成本的前提下,有效提升多分类准确率与鲁棒性。Cheng等[34]采用残差神经网络作为深层特征提取器,并结合edRVFL(ensemble deep RVFL)生成的浅层随机特征,通过集成多样化的edRVFL和两阶段超参数搜索,相比端到端深度网络显著降低了模型训练的计算开销。虽然RVFL在时间序列分类领域已经取得了长足发展,但在数据不足条件下如何有效训练RVFL、减轻性能退化,仍是一个待研究的问题。Wang等[35]通过联合域适应将标记与未标记数据映射到共享子空间,再由RVFL结合图正则化实现半监督训练,显著降低了模型训练对标记数据的依赖。叶璇等[36]基于贝叶斯理论和属性加权矩阵构建更优加权子空间,并在该子空间上训练集成RVFL模型,在高维小样本场景中表现出更好的泛化能力。Peng等[37]提出一种半监督联合优化框架,通过交替更新输出权重和未标记样本的标签矩阵,使RVFL和极限学习机能端到端协同利用标注与未标注数据,提升在标注数据稀缺场景下的分类性能。现有的研究围绕RVFL引入正则化、半监督学习和结构扩展等方法,在标注数据稀缺场景下取得了较好的分类性能。然而,这些方法大多从“整体数据集”的角度出发构建通用时间序列分类模型,在样本层面近似地将各条序列视为独立同分布的个体,未能充分利用同一周期性时间序列内部各周期间的紧密联系。
基于上述分析,为解决周期性时序数据中局部数据不足的问题,提出基于输出迁移的随机向量函数链网络(output-based transfer learning for random vector functional-link net,OTL-RVFL)。其基本思想是以周期为单位对原始时间序列进行划分,首先在数据充足的周期段上训练标准RVFL分类器并获得其未归一化的输出层预测值;当后续周期出现样本不足时,沿用与前序RVFL相同的隐藏层构建OTL-RVFL,将同一序列历史周期的RVFL预测输出与当前周期的真实标签进行加权融合,从而形成包含历史知识与当前监督信息的联合优化目标,使基于输出的迁移学习能够在序列级实现更有针对性的知识传递;最后,OTL-RVFL对该新目标仍采用RVFL的解析解求解方式计算输出层权重。基于输出的迁移方式使所提出的模型在完全保留RVFL简洁结构和解析解求解优势的基础上,能显著提升泛化能力,有效缓解数据不足所导致的分类性能下降;相较其他时序数据分类模型,OTL-RVFL依托可解析解计算输出层权重的优化机制,在训练效率和计算成本方面具有明显优势,更适合部署于对响应速度要求较高的实时应用场景。提出的模型面向周期性时序的周期段开展迁移学习,对于任意目标样本,其迁移源域一方面受整体源域集合的影响,另一方面还叠加了来自同一时间序列内部周期间的一致性约束,从而使迁移更契合目标样本的特征分布。为验证OTL-RVFL在资源受限场景中的适用性和有效性,将该模型在5个公开数据集上进行数据不足的模拟实验,并期望在保持与标准RVFL近似计算成本的前提下,提升该模型的分类准确率。
PM2.5DFCC(PM2.5 data of five Chinese cities)[44]来自UCI 机器学习库,原始数据涵盖了北京、成都、广州、上海和沈阳五座城市于2010—2015年间每小时的空气质量与气象监测记录。实验取PM2.5、露点、湿度、气压和温度五个具备周期性的指标构建实验数据集。首先,按城市和时间将数据重新分组和排序;随后,以144 h为单元将连续观测值切分为独立样本。在样本构建过程中,剔除了异常值和缺失率超过30%的样本。最终数据集以5种指标为分类类别,每类筛选出256个样本。
BMSAQ(Beijing multi-site air quality)[45]原始数据来自UCI机器学习数据库,包含北京多个空气质量监测站点的监测数据。实验数据剔除了稀疏性较高的降雨量以及非数值型风向,保留其余10个监测指标作为研究对象;接着,采用长度为 144 的非重叠时间窗口对数据进行切分,并剔除缺失率超过 30%的样本。最后,从每个指标中各筛选出256 个样本,构建了一个以指标类型为类别标签的单变量时间序列分类数据集。
WangWeina, LiMingli. Multivariate time series classification method based on Shapelets[J]. Science Technology and Engineering, 2025, 25(1): 252-261.
[3]
MiltiadousA, GionanidisE, TzimourtaK D, et al. DICE-Net: a novel convolution-transformer architecture for Alzheimer detection in EEG signals[J]. IEEE Access, 2023, 11: 71840-71858.
[4]
LinesJ, BagnallA, Caiger-SmithP, et al. Classification of household devices by electricity usage profiles[C]// Intelligent Data Engineering and Automated Learning-IDEAL 2011. Berlin: Springer, 2011: 403-412.
[5]
LiuC, WangX, TanJ, et al. Discharge voltage time series classification of lithium-ion cells based on deep neural networks[C]// 2018 IEEE 4th International Conference on Computer and Communications (ICCC). Chengdu: IEEE, 2018: 2128-2132.
[6]
WuH, HuT, LiuY, et al. TimesNet: temporal 2D-variation modeling for general time series analysis[C]// The Eleventh International Conference on Learning Representations. Kigali: ICLR, 2023: 2210.
[7]
BagnallA, LinesJ, BostromA, et al. The great time series classification bake off: a review and experimental evaluation of recent algorithmic advances[J]. Data Mining and Knowledge Discovery, 2017, 31(3): 606-660.
[8]
SakoeH, ChibaS. Dynamic programming algorithm optimization for spoken word recognition[J]. IEEE Transactions on Acoustics, Speech, and Signal Processing, 1978, 26(1): 43-49.
[9]
BerndtD J, CliffordJ. Using dynamic time warping to find patterns in time series[C]// Proceedings of the 3rd International Conference on Knowledge Discovery and Data Mining. Seattle: AAAI Press, 1994: 359-370.
ZhangYe, HouYi, OuyangKewei, et al. Survey of univariate sequence data classification methods[J]. Systems Engineering and Electronics, 2023, 45(2): 313-335.
[12]
CorduasM, PiccoloD. Time series clustering and classification by the autoregressive metric[J]. Computational Statistics & Data Analysis, 2008, 52(4): 1860-1872.
LiChen, WanYuan. Study on time series shapelets extraction based on optimization and two-phase filtering[J]. Computer Science, 2023, 50(2): 146-157.
[17]
LeX M, TranM T, HuynhV N. Learning perceptual position-aware shapelets for time series classification[C]// Machine Learning and Knowledge Discovery in Databases. Cham: Springer Nature Switzerland, 2023: 53-69.
[18]
LinesJ, TaylorS, BagnallA. Time series classification with HIVE-COTE: the hierarchical vote collective of transformation-based ensembles[J]. ACM Transactions on Knowledge Discovery from Data, 2018, 12(5): 52.
[19]
MiddlehurstM, LargeJ, FlynnM, et al. HIVE-COTE 2. 0: a new meta ensemble for time series classification[J]. Machine Learning, 2021, 110(11/12): 3211-3243.
LiKewen, KeCuihong, ZhangMin, et al. Time series classification method with local attention enhancement[J]. Computer Engineering and Applications, 2024, 60(1): 189-197.
LiuYing, LiXu, LüZheng, et al. Time series classification using deep residual probability random forests[J]. Control and Decision, 2024, 39(7): 2315-2324.
[24]
MuY, ShahzadM, ZhuX X. MPTSNet: integrating multiscale periodic local patterns and global dependencies for multivariate time series classification[C]// Proceedings of the 39th Annual AAAI Conference on Artificial Intelligence. Washington: AAAI Press, 2025: 19572-19580.
[25]
PaoY H, TakefujiY. Functional-link net computing: theory, system architecture, and functionalities[J]. Computer, 1992, 25(5): 76-79.
[26]
PaoY H, ParkG H, SobajicD J. Learning and generalization characteristics of the random vector functional-link net[J]. Neurocomputing, 1994, 6(2): 163-180.
[27]
SuganthanP N, KatuwalR. On the origins of randomization-based feedforward neural networks[J]. Applied Soft Computing, 2021, 105: 107239.
[28]
SchmidtW F, KraaijveldM A, DuinR P W. Feedforward neural networks with random weights[C]// Proceedings of 11th IAPR International Conference on Pattern Recognition. Conference B:Pattern Recognition Methodology and Systems. Hague: IEEE, 1992: 1-4.
[29]
IgelnikB, PaoY H. Stochastic choice of basis functions in adaptive function approximation and the functional-link net[J]. IEEE Transactions on Neural Networks, 1995, 6(6): 1320-1329.
[30]
HuangG B, ChenL, SiewC K. Universal approximation using incremental constructive feedforward networks with random hidden nodes[J]. IEEE Transactions on Neural Networks, 2006, 17(4): 879-892.
[31]
RenY, SuganthanP N, SrikanthN, et al. Random vector functional link network for short-term electricity load demand forecasting[J]. Information Sciences, 2016, 367/368: 1078-1093.
[32]
HuM, GaoR, SuganthanP N, et al. Automated layer-wise solution for ensemble deep randomized feed-forward neural network[J]. Neurocomputing, 2022, 514: 137-147.
[33]
HuM, HerngC J, SuganthanP N, et al. Ensemble deep random vector functional link neural network for regression[J]. IEEE Transactions on Systems, Man, and Cybernetics: Systems, 2023, 53(5): 2604-2615.
[34]
HuM, SuganthanP N. Representation learning using deep random vector functional link networks for clustering[J]. Pattern Recognition, 2022, 129: 108744.
[35]
PengT, SongS, SuoL, et al. Research and application of a novel graph convolutional RVFL and evolutionary equilibrium optimizer algorithm considering spatial factors in ultra-short-term solar power prediction[J]. Energy, 2024, 308: 132928.
[36]
HazarikaB B, GuptaD. Modelling and forecasting of COVID-19 spread using wavelet-coupled random vector functional link networks[J]. Applied Soft Computing, 2020, 96: 106626.
[37]
MalikA K, GaoR, GanaieM A, et al. Random vector functional link network: Recent developments, applications, and future directions[J]. Applied Soft Computing, 2023, 143: 110377.
MaoQi, WangYilong, WangWeijie, et al. A hybrid data-driven fault diagnosis method for traction rectifier[J]. Journal of Railway Science and Engineering, 2025, 22(8): 3795-3808.
[40]
MalikA K, GanaieM A, TanveerM, et al. Extended features based random vector functional link network for classification problem[J]. IEEE Transactions on Computational Social Systems, 2024, 11(4): 4744-4753.
[41]
ChengW X, SuganthanP N, KatuwalR. Time series classification using diversified ensemble deep random vector functional link and resnet features[J]. Applied Soft Computing, 2021, 112: 107826.
[42]
WangW, PengY, KongW. EEG-based emotion recognition via joint domain adaptation and semi-supervised RVFL network[C]// Advances in Intelligent Automation and Soft Computing. Cham: Springer, 2022: 413-422.
YeXuan, HeYulin, ZhangManjing, et al. Random vector function link network ensemble learning method for weighted subspaces[J]. Journal of Southwest China Normal University (Natural Science Edition), 2022, 47(12): 1-10.
[45]
PengY, LiQ, KongW, et al. A joint optimization framework to semi-supervised RVFL and ELM networks for efficient data classification[J]. Applied Soft Computing, 2020, 97: 106756.
[46]
WeissK, KhoshgoftaarT M, WangD. A survey of transfer learning[J]. Journal of Big Data, 2016, 3(1): 9.
[47]
WangG, ZhangG, ChoiK S, et al. An output-based knowledge transfer approach and its application in bladder cancer prediction[C]// 2017 International Joint Conference on Neural Networks (IJCNN). Anchorage: IEEE, 2017: 356-363.
SunQiao, LingWeixin. Source domain selection in transfer learning based on domain similarity rank[J]. Science Technology and Engineering, 2020, 20(20): 8245-8251.
LongMaosen, WangShitong. Feature-augmented random vector functional-link neural network[J]. Journal of Software, 2024, 35(6): 2903-2922.
[52]
ZhouH, ZhangS, PengJ, et al. Informer: beyond efficient transformer for long sequence time-series forecasting[C]// The Thirty-Fifth AAAI Conference on Artificial Intelligence. Palo Alto: AAAI Press, 2021: 11106-11115.
[53]
LaiG, ChangW C, YangY, et al. Modeling long-and short-term temporal patterns with deep neural networks[C]// Proceedings of ACM Conference (SIGIR’18). New York: ACM, 2018: 95-104.
[54]
LiangX, LiS, ZhangS, et al. PM2.5 data reliability, consistency, and air quality assessment in five Chinese cities[J]. Journal of Geophysical Research: Atmospheres, 2016, 121(17): 220-236.
[55]
ZhangS, GuoB, DongA, et al. Cautionary tales on air-quality improvement in Beijing[J]. Proceedings of the Royal Society A: Mathematical, Physical and Engineering Sciences, 2017, 473(2205): 20170457.