【作者】 陈曦；

【导师】 金秀； 袁媛；

【作者基本信息】 安徽农业大学 ， 电子信息（专业学位）， 2025， 硕士

【摘要】 时序数据作为日常生活中最常见的数据类型,在金融交易分析、医疗健康监测、工业设备故障诊断、气象预测等诸多领域发挥着关键作用。这类数据因其固有的时间依赖性和复杂的动态模式,在分析建模时带来了诸多困难。而在时序数据分类任务中,当前面临的挑战有以下几个方面:首先,大部分时序数据数据量少或难以采集,而深度模型往往需要大量训练数据提升泛化性能;其次,传统基于一维序列的处理方式难以充分捕捉数据中蕴含的多尺度空间特征。针对这些问题,本文利用维度转换将序列数据转换成二维,再通过深度生成方式进行数据增强,最终提出了由最优维度转换方法、生成方法和分类方法组成的时序数据分类框架。研究内容主要包含以下三个方面:(1)研究并比较了不同维度转换方法。实验选择了UEA时序数据分类库中的五个数据集作为研究对象,探索了四种具有代表性的时序数据维度转换方法:格拉姆角差场(GADF)、短时傅里叶变换(STFT)、马尔可夫转移场(MTF)和梅尔频谱图(MS)。通过设计对比实验和可视化分析,发现GADF方法的综合性能最佳。实验表明,经GADF转换后的数据可使LSTM分类器的平均准确率从63.5%提升至70.7%,相对提升幅度达11.3%。特别是,该方法在Basic Motions数据集上的准确率达到了30%的提升,验证了维度转换方法在时序数据集上的有效性。(2)探究了二维时序数据中不同生成方法的结果,并讨论了CycleGAN模型中不同数据源与不同生成器判别器的组合对模型性能的影响。针对数据量少及难以采集的问题,本文比较了六种数据增强方法(包括传统方法和深度生成方法),发现CycleGAN在生成质量和多样性方面具有显著优势。实验数据显示,该方法在Finger Movements和Racket Sports数据集上分别提升了4%和7.2%的分类准确率。进一步研究其生成机制发现,采用UNet＿256生成器与70×70 Patch GAN判别器的组合最适合时序数据集。此外,数据源选择对生成的效果也具有一定的影响,由另一组数据集生成的数据更加适合增强当前数据集。(3)基于前两个阶段的研究成果,对比了不同分类器的性能,并提出了名为CycleTime的分类框架。实验对比了CNN、RNN、LSTM及FCN四种分类器,FCN在平均准确率、召回率和精确度上分别提高了2.4%、2.4%和2.2%,优于其他分类器。并且在Epilepsy数据集上的准确率达到0.833。因此,CycleTime框架由GADF、CycleGAN与FCN组成。最后,实验将CycleTime与其他六种常见的分类方法做对比,其平均分类准确率达到75.1%,并且其各项评价指标都为最佳。特别是在Epilepsy数据集上,该框架的准确率为83.3%,召回率和精确度分别达到83.8%和84.1%,显著优于其他分类方法。验证了该框架的有效性。综上所述,本文针对时序数据分类任务所提出的方法能够有效解决时序数据数据量少的问题,并显著提升了分类模型的各项性能。其研究结果对于时序数据分析的各项任务都具有重要的参考价值。更多还原

【Abstract】 Time series data,as one of the most common data types in daily life,plays a crucial role in various fields such as financial transaction analysis,healthcare monitoring,industrial equipment fault diagnosis,and weather forecasting.Due to its inherent temporal dependencies and complex dynamic patterns,analyzing and modeling such data present numerous challenges.In time series classification tasks,the current challenges include the following aspects:First,most time series datasets are small or difficult to collect,while deep models often require large amounts of training data to improve generalization performance.Second,traditional one-dimensional sequence-based processing methods struggle to fully capture the multi-scale spatial features inherent in the data.To address these issues,this thesis employs dimension transformation to convert sequential data into two-dimensional representations,followed by deep generative methods for data augmentation,ultimately proposing a time series classification framework consisting of optimal dimension transformation,generation,and classification methods.The main research contents include the following three aspects:(1)Investigation and comparison of different dimension transformation methods.This study selects five datasets from the UEA Time Series Classification Repository as research subjects and explores four representative time series dimension transformation methods:Gramian Angular Difference Field(GADF),Short-Time Fourier Transform(STFT),Markov Transition Field(MTF),and Mel Spectrogram(MS).Through comparative experiments and visualization analysis,GADF is found to achieve the best overall performance.Experiments show that data transformed by GADF can improve the average accuracy of an LSTM classifier from 63.5%to 70.7%,representing a relative improvement of 11.3%.Notably,the accuracy on the Basic Motions dataset increases by 30%,verifying the effectiveness of dimension transformation methods on time series datasets.(2)Exploration of different generation methods for two-dimensional time series data and discussion on the impact of different data sources and generator-discriminator combinations in the CycleGAN model on model performance.To address the challenges of small and difficult-to-collect datasets,this study compares six data augmentation methods,including traditional methods and deep generative methods,and finds that CycleGAN has significant advantages in generation quality and diversity.Experimental results show that this method improves classification accuracy by 4%and7.2%on the Finger Movements and Racket Sports datasets,respectively.Further investigation into its generative mechanism reveals that the combination of a UNet＿256generator and a 70×70 Patch GAN discriminator is the most suitable for time series datasets.Additionally,data source selection has a certain impact on generation effectiveness,with data generated from an alternative dataset proving more effective in enhancing the current dataset.(3)Comparison of different classifiers’performance and proposal of the CycleTime classification framework.Based on the findings from the previous two stages,this study compares the performance of CNN,RNN,LSTM,and FCN classifiers.The results show that FCN outperforms the other classifiers,with improvements of 2.4%,2.4%,and 2.2%in average accuracy,recall,and precision,respectively.Furthermore,it achieves an accuracy of 0.833 on the Epilepsy dataset.As a result,the CycleTime framework consists of GADF,CycleGAN,and FCN.Finally,experiments compare CycleTime with six other commonly used classification methods,achieving an average classification accuracy of 75.1%and the best performance across all evaluation metrics.Notably,on the Epilepsy dataset,this framework achieves an accuracy of 83.3%,with recall and precision reaching 83.8%and 84.1%,respectively,significantly outperforming other classification methods and verifying the effectiveness of the framework.In conclusion,the method proposed in this thesis effectively addresses the problem of limited time series data and significantly enhances the performance of classification models.The research findings hold important reference value for various time series analysis tasks.更多还原