【作者】 陈博；

【导师】 方敏；

【作者基本信息】 西安电子科技大学 ， 计算机科学与技术， 2024， 博士

【摘要】 随着各种传感器采集设备的不断发展,数据量随着时间的推移急剧增加,时间序列分析成为热点研究内容,包括但不限于时间序列的分类与异常检测。传统的时间序列分类方法需要考虑整体时间序列的时序性语义。然而,在工业应用中,随着时间推移所采集的数据时间维度,即时间序列数据点数量逐渐提升,整体序列时序性语义的计算复杂度也持续提升。幸运的是,由于时间序列数据按照固定采样频率采集,其冗余数据导致时间序列中包含许多相似或重复的信息,因此产生了基于局部判别性子序列(shapelet)的分类方法,以显著降低计算复杂度。在该方法框架中,每个时间序列可以由一定数量的shapelet表示,避免整体序列计算。因此,如何在时间序列中高效发现这些判别性子序列并解释其分类结果成为当前机器学习和时间序列分类的研究热点,也是本文的研究重点。此外,由于时间序列维度随采集过程维度逐渐升高,难以捕捉时间依赖,导致序列异常事件相对于整体时间序列具有稀疏性与不可预测性,且异常的表现形式多样,难以标注。因此,如何在保证时间依赖的前提下检测时间序列中的异常事件也是本文研究重点。基于shapelet的时间序列分类方法往往受限于海量子序列的挖掘或学习过程。因此,本文提出了一系列方法,旨在提升基于shapelet的时间序列分类算法性能,并结合判别性子序列的局部模式,在海量交通流数据中检测稀疏异常事件。首先,假设时间序列数据同时具有特征重要性与时间重要性,以发现重要时间点,并利用这些重要时间点来匹配最优的shapelet。其次,从语义角度分析shapelet的潜在分布,学习样本特定的shapelet,用于时间序列分类。再次,结合无监督聚类算法的优点,将shapelet学习过程视为连续时间点的无监督时间聚类过程,以辅助shapelet的学习。最后,借助关键子序列的局部模式来探索检测交通流数据中稀疏异常事件,以提升时间序列异常检测性能。具体而言,本文的工作围绕时间序列维度约简、时间依赖、时间序列异常检测以及基于shapelet的时间序列分类算法展开,主要贡献如下:1.针对基于学习shapelet的时间序列分类算法中判别性子序列形状任意与泛化性低的问题,提出了一种基于特征选择的噪声增强shapelet选择模型,模型通过特征选择策略动态选择最相关的噪声增强shapelet进行时间序列分类。首先,该模型使用分类模型生成实例级特征重要度评分。同时,基于循环神经网络的模型计算时域重要度评分。最后,模型使用这两个重要度评分集成为软标签,以指导模型选择样本特定shapelet,从而选择最具判别性的shapelet进行时间序列分类。实验部分表明,模型在22个UCR数据集上分类准确度优于8种先进对比方法,并在3个数据集上达到最优分类。2.针对基于学习shapelet的时间序列分类算法中判别性子序列语义模糊与潜在分布难以估计问题,本文提出了一种shapelet解耦生成模型,学习样本特定的shapelet用于时间序列分类。该模型使用序列变分自编码器解耦shapelet的三个生成因子,即动态时变、动态时不变因子和静态因子,这些样本特定的shapelet可以根据片段级的不同输入生成。此外,模型利用可微分动态时间规整指导样本生成,且通过动态时变因子和动态时不变因子建模的条件分布进行采样,确保生成的样本与相应的子序列相似且同时具有多样性。同时,模型采用成对损失约束动态时不变因子,使序列的波动保持在规定的时间邻域内,从而缓解序列相移和失真问题。最后,由于动态因子和动态时不变因子都对应于原始序列的每个子序列,模型通过卷积运算学习每个序列的样本特定shapelet,形成具有可解释分类结果的判别性特征,实现高效分类。实验部分表明,模型在20个UCR数据集上分类准确度优于8种先进对比方法,并在5个数据集上达到最优分类效果。3.针对基于搜索shapelet的时间序列分类算法中子序列候选与原样本相关性丢失问题,本文提出了一种基于时域聚类的shapelet学习算法用于时间序列分类。该算法采用非负矩阵三因子分解和时间正则化方法对时间序列时域维度特征进行聚类。该过程在时间序列中寻找显著的连续时间点子集,避免穷举搜索或构建大量子序列样本,有效地降低了聚类复杂度,并为shapelet优化提供了有效范围。同时,模型使用时域特征质心指导shapelet优化,并结合聚类和样本标签之间的相关性构建shapelet正则化。由时域特征质心引导的优化确保了 shapelet与原始序列相似,且同时避免shapelet成为任意形状。最后,模型使用聚类和标签之间的相关性指导shapelet转换表示矩阵,进一步增强shapelet优化,产生可判别低维表示,提高分类性能。实验部分表明,模型在20个UCR数据集上分类准确度优于6种对比方法,并在8个数据集上优于最先进的时间序列分类算法。4.针对交通流时间序列数据实际场景中异常事件稀疏、路网模式复杂以及标注成本高的问题,本文提出了一种基于判别子序列泛化的交通流异常检测模型。该模型通过先验异常模式融合与对比学习表示,有效的将先验知识以同一度量标准融入到模型中,辅助异常检测。同时,模型使用异常特定方向的噪音进行特征增强,使得在训练数据与测试数据存在统计性差异时,模型可以将先验知识鲁棒的泛化在未见路段中,提高异常检测性能。实验部分表明,模型在真实交通数据中相比于先进异常检测算法精确率提升26%,召回率提升18%,实现高效检测异常,并表现出较强泛化性。更多还原

【Abstract】 With the rapid development of various sensor devices,the volume of data collected over time has dramatically increased,making time series analysis a key research focus,encompassing but not limited to time series classification and anomaly detection.Traditional time series classification methods often require consideration of the temporal semantics of the entire series.However,in industrial applications,as data is continuously collected over time,the temporal dimension increases,leading to a continuous rise in the computational complexity of capturing the overall temporal semantics.Fortunately,because time series data is typically collected at a fixed sampling rate,there is often a large amount of redundant data within the series.This has led to the development of shapelet-based classification methods that focus on locally distinguishable subsequences(shapelets),significantly reducing computational complexity.In this framework,each time series can be represented by a certain number of shapelets,avoiding the need for full-sequence computation.Efficiently identifying these distinguishable subsequences in time series has become a major research focus in machine learning and time series classification.Moreover,as the dimensionality of time series increases with ongoing data collection,capturing temporal dependencies becomes more challenging,leading to the sparsity and unpredictability of anomalous events,which can manifest in various forms and is difficult to be labeled.Therefore,detecting anomalies in time series while preserving temporal dependencies is another key focus of this study.Shapelet-based time series classification methods are often constrained by the challenges of mining or learning from massive subsequences.Therefore,this study proposes a series of methods aimed at enhancing the performance of shapelet-based time series classification algorithms and detecting sparse anomalous events in large-scale traffic flow data by incorporating the local patterns of key subsequences.First,we assume that time series data exhibit both feature importance and temporal importance,enabling the identification of critical time points that can be used to match optimal shapelets.Second,we analyze the potential distribution of shapelets from a semantic perspective,generating sample-specific shapelets for improved classification.Third,leveraging the advantages of unsupervised clustering algorithms,we frame the shapelet learning process as an unsupervised temporal clustering process over continuous time points to facilitate shapelet learning.Finally,we utilize the local patterns of key subsequences to explore and detect sparse anomalous events within traffic flow data,thereby enhancing the performance of time series anomaly detection.Specifically,this work focuses on dimensionality reduction in time series,temporal dependencies,anomaly detection,and the development of shapelet-based time series classification algorithms,with the following main contributions:1.To address the issues of arbitrary shapes and low generalization in discriminative subsequences within shapelet-based time series classification algorithms,this study introduces a feature selection-based noise-enhanced shapelet selection model.This model dynamically selects the most relevant noise-enhanced shapelets for time series classification through a feature selection strategy.First,the model generates instance-level feature importance scores using a classification model,while simultaneously computing temporal importance scores based on a recurrent neural network.These two sets of importance scores are then combined into soft labels to guide the model in selecting sample-specific shapelets,ultimately identifying the most discriminative shapelets for time series classification.Experimental results demonstrate that the model outperforms eight advanced comparative methods in classification accuracy across 22 UCR datasets,achieving optimal classification on three of these datasets.2.To address the issues of semantic ambiguity and difficulty in estimating the potential distribution of discriminative subsequences in shapelet-based time series classification algorithms,this study proposes a shapelet disentangling learning model based on generative networks.This model is designed to learn sample-specific shapelet and time series classification.The model utilizes a sequence variational autoencoder to decouple three generative factors of shapelets:dynamic time-varying factors,dynamic time-invariant factors,and static factors.These sample-specific shapelets can be generated based on segment-level inputs.Additionally,the model employs differentiable dynamic time warping to guide sample generation and samples from the conditional distribution modeled by the dynamic and time-invariant factors,ensuring that the generated samples are similar to their corresponding subsequences while maintaining diversity.Moreover,the model incorporates a pairwise loss constraint on the dynamic time-invariant factors,keeping the fluctuations of the sequences within a specified temporal neighborhood to mitigate issues of sequence shifting and distortion.Finally,since both dynamic factors and dynamic time-invariant factors correspond to each subsequence of the original series,the model learns sample-specific shapelets for each sequence through convolution operations,resulting in discriminative features that yield interpretable classification results for efficient classification.Experimental results indicate that the model outperforms eight advanced comparative methods in classification accuracy across 20 UCR datasets,achieving optimal classification on five of these datasets.3.To address the issue of losing the correlation between candidate subsequences and the original samples in search-based shapelet time series classification algorithms,this study proposes a discriminative shapelet learning algorithm based on temporal clustering and matrix factorization for time series classification.This algorithm employs non-negative matrix tri-factorization and temporal regularization methods to cluster the temporal domain features of time series.The process identifies significant subsets of continuous time points within the time series,effectively reducing clustering complexity by avoiding exhaustive searches or the construction of numerous subsequence samples,thus providing an effective range for shapelet optimization.Additionally,the model uses the centroids of temporal features to guide shapelet optimization and constructs shapelet regularization by considering the correlation between clustering and sample labels.This optimization,directed by the centroids,ensures that shapelets remain similar to the original sequences while preventing them from becoming arbitrary shapes.Finally,the model enhances shapelet optimization by utilizing the correlation between clustering and labels to guide the shapelet transformation representation matrix,resulting in a discernible low-dimensional representation that improves classification performance.4.To address the challenges of sparse anomalous events,complex road network patterns,and high labeling costs in real-world traffic flow time series data,this study proposes a traffic flow anomaly detection model that integrates discriminative subsequence with domain generalization.This model effectively incorporates prior knowledge through a fusion of prior anomalous patterns and contrastive learning representations,aligning these elements under a unified metric to assist in anomaly detection.Additionally,the model enhances features using noise specific to anomalous directions,enabling robust generalization of prior knowledge in unseen segments when statistical differences exist between training and testing data.This approach improves the model’s performance in detecting anomalies.The model outperforms the baseline methods in both precision and recall on real traffic data,while achieving higher recall on synthetic data compared to the baselines.The experimental results demonstrate that the model improves precision by 26%and recall by 18%compared to state-of-the-art anomaly detection algorithms on real traffic data.This highlights its capability for efficient anomaly detection and strong generalization performance.更多还原