Method of Time Series Similarity Measurement Based on Dynamic Time Warping

With the rapid development of mobile communication all over the world,the similarity of mobile phone communication data has received widely attention due to its advantage for the construction of smart cities.Mobile phone communication data can be regarded as a type of time series and dynamic time warping(DTW)and derivative dynamic time warping(DDTW)are usually used to analyze the similarity of these data.However,many traditional methods only calculate the distance between time series while neglecting the shape characteristics of time series.In this paper,a novel hybrid method based on the combination of dynamic time warping and derivative dynamic time warping is proposed.The new method considers not only the distance between time series,but also the shape characteristics of time series.We demonstrated that our method can outperform DTW and DDTW through extensive experiments with respect to cophenetic correlation.

Partition of GB-InSAR deformation map based on dynamic time warping and k-means

Ground-based interferometric synthetic aperture radar(GB-InSAR)can take deformation measurement with a high accuracy.Partition of the GB-InSAR deformation map benefits analyzing the deformation state of the monitoring scene better.Existing partition methods rely on labelled datasets or single deformation feature,and they cannot be effectively utilized in GBInSAR applications.This paper proposes an improved partition method of the GB-InSAR deformation map based on dynamic time warping(DTW)and k-means.The DTW similarities between a reference point and all the measurement points are calculated based on their time-series deformations.Then the DTW similarity and cumulative deformation are taken as two partition features.With the k-means algorithm and the score based on multi evaluation indexes,a deformation map can be partitioned into an appropriate number of classes.Experimental datasets of West Copper Mine are processed to validate the effectiveness of the proposed method,whose measurement points are divided into seven classes with a score of 0.3151.

Grey incidence clustering method based on multidimensional dynamic time warping distance

The traditional grey incidence degree is mainly based on the distance analysis methods, which is measured by the displacement difference between corresponding points between sequences. When some data of sequences are missing (inconsistency in the length of the sequences), the only way is to delete the longer sequences or to fill the shorter sequences. Therefore, some uncertainty is introduced. To solve this problem, by introducing three-dimensional grey incidence degree (3D-GID), a novel GID based on the multidimensional dynamic time warping distance (MDDTW distance-GID) is proposed. On the basis of it, the corresponding grey incidence clustering (MDDTW distance-GIC) method is constructed. It not only has the simpler computation process, but also can be applied to the incidence comparison between uncertain multidimensional sequences directly. The experiment shows that MDDTW distance-GIC is more accurate when dealing with the uncertain sequences. Compared with the traditional GIC method, the precision of the MDDTW distance-GIC method has increased nearly 30%.

Hand Gesture Recognition by Accelerometer-Based Cluster Dynamic Time Warping

Aiming at the diversity of hand gesture traces by different people,the article presents novel method called cluster dynamic time warping( CDTW),which is based on the main axis classification and sample clustering of individuals. This method shows good performance on reducing the complexity of recognition and strong robustness of individuals. Data acquisition is implemented on a triaxial accelerometer with 100 Hz sampling frequency. A database of 2400 traces was created by ten subjects for the system testing and evaluation. The overall accuracy was found to be 98. 84% for user independent gesture recognition and 96. 7% for user dependent gesture recognition,higher than dynamic time warping( DTW),derivative DTW( DDTW) and piecewise DTW( PDTW) methods.Computation cost of CDTW in this project has been reduced 11 520 times compared with DTW.

Efficient Dynamic Time Warping by Adaptively Controlling the Valid Warping Range

Dynamic time warping(DTW)spends most of the time in generating the correlation table,and it establishes the global path constraints to reduce the time complexity.However,the global constraints restrain just in terms of the time axis.In this paper,we therefore propose another version of DTW,to be called branch-and-bound DTW(BnB-DTW),which adaptively controb its global path constraints by reflecting the contents of input patterns. Experimental results show that the suggested BnB-DTW algorithm performs more efficiently than other conventional DTW approaches while not increasing the optimal warping cost.

基于PCA-ShapeDTW-QWGRU的分布式光伏集群短期功率预测

针对分布式光伏短期功率预测建立基于主成分分析、改进的动态时间规整算法与量子加权门控循环单元(PCAShapeDTW-QWGRU)的集群功率预测模型。针对集群划分不够精细、光伏电站数据蕴含的信息难以捕捉的问题,提出基于主成分分析结合密度聚类算法(PCA-OPTICS)的集群划分方法;针对目前选取代表电站与集群相似性较低的问题,提出基于改进的动态时间规整算法(ShapeDTW)的代表电站的选取方法,利用ShapeDTW度量相似性距离,选取最小值作为代表电站,并利用基于均方根传播梯度下降法优化的量子加权门控循环单元(RMSprop-QWGRU)模型进行预测;为了解决代表电站与集群功率的变换系数转换差异较大的问题,采用实时变换系数对代表电站进行集群功率值预测计算。实验结果表明,所提方法能有效提升光伏集群功率预测的精度。

基于多尺度LDTW和TCN的空间负荷预测方法

空间负荷预测为合理建设和使用变电站、馈线等提供了重要的指导,成为配电网规划中不可或缺的一部分。配电网规划的精细化产生了大量高分辨率的负荷数据,社会的快速发展使得地块的用电特征日趋复杂。当前的空间负荷预测没有充分考虑负荷数据之间的时间特性,且在预测过程中也未考虑到不同类型地块间可能存在的负荷峰值出现时间不一致问题。为此,提出一种空间负荷预测方法,通过基于多尺度限制对齐路径长度(LDTW)的谱聚类分析用户的负荷曲线在形状上的相似性,并提取不同地块的典型用电行为,以进一步分类确定同类型地块对应的同时率。多尺度LDTW通过限制序列之间匹配步长的上限来抑制病态匹配的产生,提高曲线相似性的综合评估能力。根据聚类结果筛选适合待预测区域的训练样本并构建基于时间卷积网络(TCN)的回归预测模型,将预测结果基于地块各自的同时率进行聚合,实现空间负荷预测。实验结果表明:该方法加强了对负荷曲线形状的分析和对不同类型地块同时率的区分,在聚类方面,DBI指数达到0.57,VI指数达到0.31;在预测方面,相对误差达到1.93%,决定系数达到0.941,相比其他典型方法均取得了较大改善。

基于DTW时序距离的故障可诊断性评估方法

对装备中具体故障进行可诊断性评估,可以量化地表明故障被诊断的难易程度。针对故障可诊断性评价方法中存在的不客观及度量方法选择不恰当的问题,提出一种基于动态时间规整(Dynamic Time Warping,DTW)的时序距离度量方法。定义可诊断性评估模型,从实际系统中抽取系统的结构、信号、测试、故障模式4种要素,为故障可诊断性评估工作的开展打下基础。摒弃传统的从信号中提取特征进行相似性度量的做法,将信号看做时序序列,基于DTW方法衡量不同状态下各信号之间的相似程度,作为故障可诊断性的评估依据。使用动量轮系统对新方法进行仿真实验,仿真结果表明新方法在进行可诊断性评估方面具有客观性和有效性。

基于改进型DTW的语音识别系统的设计与实现

大部分系统使用深度学习技术完成语音识别任务,并取得良好的效果。但是,基于深度学习的语音识别技术对计算机硬件算力的要求较高,同时需要大量的语音样本对语音识别模型进行训练。针对这些问题,基于改进后的动态时间归整(Dynamic Time Warping,DTW)算法设计并实现了一个语音识别系统。

改进DTW下界约束的Granger多元时序LSTM预测模型

多元时序的因果预测研究是探讨复杂网络响应关系的热点问题。提出一种通过DTW的下界约束组合并改进的层级过滤器,与格兰杰因果验证法相结合验证其因果统计量,挖掘出有效信息实现有效降维,进一步输入LSTM预测模型进行因果预测。仿真实验利用开源的空气质量数据集进行定量和定性对比验证,该方法的损失函数训练曲线和测试曲线有较好的拟合度,表明该因果预测法是可行且有效的。

基于DTW-KiBaM模型的锂电池SOC估计

锂离子电池工作温度及老化程度是影响其荷电状态(State of Charge,SOC)估计算法准确性的关键因素.在二阶RC等效电路模型的基础上,提出了一种由动态时间规整算法(Dynamic Time Warping,DTW)与KiBaM模型(Kinetic Battery Model,KiBaM)相结合的混合模型.利用DTW算法基于充电电压数据确定电池老化状态,通过KiBaM模型计算电池由于电流效应导致的不可用容量,结合二阶RC等效电路模型推导新的SOC计算矩阵,采用无迹卡尔曼滤波算法(Unscented Kalman Filter,UKF)实现SOC估计.基于城市道路循环工况(Urban Dynamometer Driving Schedule,UDDS)验证了混合模型的准确性,实验结果表明,在10℃低温环境或经过200次循环老化后工作的锂离子电池,模型估计SOC的最大误差小于2%.

DTW算法支持下的线状要素连续地图综合方法

面向线状地图要素连续尺度变换问题,本文提出了一种DTW算法支持下的连续综合方法。该方法基于尺度融合的思想,将同一地理实体在大小两种比例尺下以不同的几何表达作为输入,首先基于DTW算法建立两种几何表达坐标顶点之间的对应关系;然后采用线性内插方法动态派生任意中间尺度上几何数据,从而实现连续地图综合。顶点之间对应关系的正确性,直接决定了线性内插的结果,而同一实体在不同比例尺下的几何表达往往具有不同的坐标点数,顶点之间具有一对多的对应关系。为寻求最优顶点匹配方案,以顶点距离作为匹配代价,以整体最小距离作为目标函数,采用DTW算法求解最优匹配。试验结果表明,基于DTW的顶点匹配方法可适应不同的地图综合场景,该方法支持下的地图综合效果可实现连续、光滑的渐变,符合地图表达规则和人类空间认知。

DTW算法在地震时间序列信息挖掘中的应用

采用DTW算法用于实现地震时间序列“异常形态”的回溯性检验与统计,以此提高地震异常信息的深度挖掘效率。随机选取太原地震监测中心站(以下简称太原站)水平摆观测的一段数据作为测试数据,验证该方法的可行性,并利用昔阳地震监测站(以下简称昔阳站)水平摆实际观测数据,自动识别出自观测以来受降雨影响的异常形态。在此基础上,将最近一次数据变化作为模板,进行回溯性检验,共提取出7次类似变化,与之前总结的异常特征一致。经分析,其中4次异常之后对应了太原盆地及其周围M3.6以上地震,运用R值评分对其进行预报效能评估,结果显示该异常形态可通过预报效能检验(R=0.52,R_(0)=0.45),其最佳预测时间为50~90 d内,可作为短期预测指标。

基于微分思想和动态时间弯曲(DTW)的飞行员着舰技能评估研究

针对舰载机着舰训练,提出了一种基于微分思想和动态时间弯曲(dynamic time warping, DTW)的飞行员驾驶技能评估方法。首先,将着舰过程分解为多元时间序列,通过对着舰过程的分析,选取飞行高度、下沉速度和对中偏差等5个特征参数作为指标,并且拟合出各指标的标准曲线;然后,对选取的特征参数进行微分分段处理,利用特征参数间的关系进行降维表示;最后,利用DTW对评估模型进行计算,实现着舰技能的评估。算例分析表明,基于微分思想和DTW方法对飞行员着舰技能评估具有很好的适用性,有利于促进飞行员着舰技能的提高。

一种基于DTW算法的磁场相似性度量方法

在舰船磁场模拟及磁探测系统的目标识别中,需要对模拟或探测磁场与真实目标磁场分布的相似程度进行评估。文中针对以往评估方法存在的问题,提出利用动态时间规整(DTW)算法计算磁场的相似度,解决了目标速度不同和起止点不一致情况下的磁场曲线相似度评估问题;为降低磁场曲线局部扰动带来的影响,利用最长公共子串对DTW算法进行了优化,提高了相似度计算精度;最后以模型试验验证了算法的正确性。与传统评估方法相比,该方法无需人工预先设置参数,能直接给出相似度评价结果,可应用于舰船磁场模拟效果评估及磁探测系统的目标识别。

基于熵值比-DTW度量奇异值相似度指标的滚动轴承退化历程辨识

针对滚动轴承健康状态监测现有退化指标单调性及鲁棒性差,且数据波动造成不同退化状态区分度低等问题,提出了一种基于熵值比-动态时间规整(DTW)度量的奇异值相似度指标并用于辨识滚动轴承不同退化历程。首先,利用奇异值分解(SVD)算法对不同时刻采集的轴承信号矩阵进行分解,将奇异值作为轴承退化特征;其次,基于DTW算法计算轴承连续退化奇异值时间序列的相似度,表征轴承的全寿命周期历程;最后,考虑到轴承不同退化状态的差异性,将熵值比作为权值对相似度指标进行优化,提高相似度退化指标的单调性及对早期异常点的敏感性。对IMS轴承全寿命周期数据的研究结果表明:奇异值相似度指标的单调性、鲁棒性及敏感性较好,可有效避免数据波动对轴承健康状态带来的干扰,能更准确地反映轴承全寿命退化历程。

基于DTW距离的普通话儿化音变差异分析

为了量化儿化音变对音节的改变量,基于动态时间规整算法,计算汉语普通话儿化音节音变前后最小对立对之内的声学距离(EX)、重复同一未发生儿化音变音节之间的距离(NIN)和重复同一儿化音节之间的距离(EIN),并将EX距离与NIN、EIN距离的差值EX-NIN、EX-EIN作为衡量儿化音变对声学距离改变量的参数。结果表明,EX距离显著大于NIN距离和EIN距离。不同发音人儿化音变改变量之间具有一定的人际差异。从不同词项、未儿化韵母和儿化韵的角度来看,均能得出拼合型儿化韵音变量总体大于融合型音变量的结论。其中,[i]-[iər]的儿化改变量显著大于其他元音。在儿化音节属于不同声调时,上声时长最长,因此儿化音变量最大;去声时长最短,所以儿化改变量最小。单音节词的儿化音变改变量显著大于双音节词的改变量。性别与框架句对儿化音变改变并没有显著影响。

High-resolution Load Profile Clustering Approach Based on Dynamic Largest Triangle Three Buckets and Multiscale Dynamic Warping Path Under Limited Warping Path Length

With the popularity of smart meters and the growing availability of high-resolution load data, the research on the dynamics of electricity consumption at finely resolved timescales has become increasingly popular. Many existing algorithms underperform when clustering load profiles contain a large number of feature points. In addition, it is difficult to accurately describe the similarity of profile shapes when load sequences have large fluctuations, leading to inaccurate clustering results. To this end, this paper proposes a high-resolution load profile clustering approach based on dynamic largest triangle three buckets(LTTBs) and multiscale dynamic time warping under limited warping path length(LDTW). Dynamic LTTB is a novel dimensionality reduction algorithm based on LTTB. New sequences are constructed by dynamically dividing the intervals of significant feature points. The extraction of fluctuation characteristics is optimized. New curves with more concentrated features will be applied to the subsequent clustering. The proposed multiscale LDTW is used to generate a similarity matrix for spectral clustering, providing a more comprehensive and flexible matching method to characterize the similarity of load profiles. Thus, the clustering effect of a high-resolution load profile is improved. The proposed approach has been applied to multiple datasets. Experiment results demonstrate that the proposed approach significantly improves the Davies-Bouldin indicator(DBI) and validity index(VI). Therefore, better similarity and accuracy can be achieved using high-resolution load profile clustering.

基于DTWKNN的电力缺失数据补全方法

电力数据是电力负荷预测工作的基础,但是由于智能电表故障、传输信道异常等原因导致的电力数据缺失现象,直接影响电力负荷预测的准确性。针对此问题,提出了一种动态时间规整k最近邻(dynamic time warping k-nearest neighbor, DTWKNN)的电力缺失数据补全方法。在传统k最近邻(KNN)方法的基础上,用动态时间规整距离取代欧式距离,参考主成分分析方法计算属性相关性影响参数,该参数调整了近邻点权重并用于补全数据,最终对所有缺失点依次补全。以某市实际用电数据进行补全实验,与KNN补全方法对比,基于DTWKNN方法的缺失数据补全效果更好。此外,经补全后的电力数据在负荷预测实验中误差更低,具备良好的电力负荷预测应用性。

基于DTW-TCN的股票分类及预测研究

随着社会以及信息技术的发展,金融工具、股票交易呈现出新的形态,其中,由于金融数据量呈指数级增长,股票类数据难以分类与预测,因此在高频交易中股票趋势预测尤为重要。为提高高频交易中股票趋势预测的精准度,构建基于动态时间规整(DTW)聚类分析的时间卷积神经网络(TCN)模型用于股票分类和预测研究。在本文模型(DTW-TCN)中,采用开盘价、最高价、最低价、收盘价、成交量、成交额作为股票特征变量。为避免量级影响,首先,对特征向量标准化处理,随后利用动态时间规整对于时间序列相似性的衡量作用,作为股票的分类标准;然后,通过TCN卷积神经网络提取类别共同特征进行网络训练,进一步,将类别中的普遍性行业股票利用训练好的卷积神经网络进行股票趋势预测;最终,得到所属类别股票每分钟开盘价与收盘价走势,并与实际趋势相对比进行误差分析。以19只行业代表性股票分钟级数据为样本进行实验,结果表明,本文模型能有效地分类趋势趋同的股票,并且实现在分钟级别高频交易中准确进行趋势预测,对比传统时间序列模型和LSTM网络模型具有更大时间特性优势。未来DTW-TCN分类预测模型可以用于更多大数据信息分类和预测实例中。