基于无监督学习的时序序列故障诊断方法研究

随着信息技术和传感器技术的发展,基于数据驱动的故障诊断技术是保障大型工业设备高效、安全运行的关键技术之一.机器学习由于其具有强大的特征表示能力与基于大数据的特征提取优势,多种深度神经网络模型已成为故障诊断领域中最常用的特征提取方法之一.但监测设备收集到的数据中包括大量的无标签数据,基于监督学习的深度神经网络模型没有办法对其进行充分的利用,造成了部分有用信息的浪费.针对无标签数据,提出基于无监督学习的方法,利用最大化互信息的思想训练特征提取模型,在此基础上,设计一种针对时序序列数据的故障诊断方法,并在公开数据集凯斯西储大学轴承数据集上验证,取得了比以往传统方法更高的诊断精度.在卫星监测数据上进一步验证,提出的特征提取模型能够对故障不同阶段进行区分,很好地捕捉不同阶段的数据特性.结果表明,提出的基于无监督学习的故障诊断方法能够有效、充分地利用大量的无标签数据,提高时序序列数据的故障诊断精度.

基于电压序列相似性的户变关系与相别识别

随着低压配电网的改造升级,台区户变关系变化频繁,为解决时有发生的用户台区挂错现象,提出一种利用改进的基于密度的点排序识别聚类结构(ordering points to identify the clustering structure,OPTICS)的台区户变关系识别和相别识别方法。首先,对配网电压序列的相关性进行定性分析,提出利用电压时序序列作为分析识别的数据基础;其次,采用改进的自适应分段聚合近似(adaptive piecewise aggregate approximation,APAA)对电压序列进行降维处理,提取能够反映电压特征的低维向量;然后利用改进的OPTICS算法对所提取的特征向量进行聚类分析,识别台区的户变关系和相别关系;最后,基于实际的台区数据进行算例分析,验证了所提方法的准确性。

基于确定性学习的人体步态识别方法研究

基于确定性学习理论,提出了一种基于关节角时序数据序列的人体步态识别方法。首先,由人体运动捕捉设备获取关节角时序数据序列,则局部准确的人体步态的内部动力学可通过径向基函数（RBF）网络得到逼近。进一步,证明了逼近误差和相关神经网络（NN）参数的收敛。接下来,通过将NN逼近得到的步态动力学知识存储于常值的RBF网络,可实现人体移动步态特征的有效表达。最后,通过构建步态模式的相似性定义,提出了一种步态时序数据识别的方法,最终可实现准确的步态识别。仿真实验采用类圆规双足机器人验证了所提方法的有效性。

A Novel Parallel Scheme for Fast Similarity Search in Large Time Series

The similarity search is one of the fundamental components in time series data mining,e.g.clustering,classification,association rules mining.Many methods have been proposed to measure the similarity between time series,including Euclidean distance,Manhattan distance,and dynamic time warping(DTW).In contrast,DTW has been suggested to allow more robust similarity measure and be able to find the optimal alignment in time series.However,due to its quadratic time and space complexity,DTW is not suitable for large time series datasets.Many improving algorithms have been proposed for DTW search in large databases,such as approximate search or exact indexed search.Unlike the previous modified algorithm,this paper presents a novel parallel scheme for fast similarity search based on DTW,which is called MRDTW(MapRedcuebased DTW).The experimental results show that our approach not only retained the original accuracy as DTW,but also greatly improved the efficiency of similarity measure in large time series.

An Improving Indexing Approach on Time Series Based on Minimum Bounding Rectangle

A fundamental problem in whole sequence matching and subsequence matching is the problem of representation of time series.In the last decade many high level representations of time series have been proposed for data mining which involve a trade-off between accuracy and compactness.In this paper the author proposes a novel time series representation called Grid Minimum Bounding Rectangle(GMBR) and based on Minimum Bounding Rectangle.In this paper,the binary idea is applied into the Minimum Bounding Rectangle.The experiments have been performed on synthetic,as well as real data sequences to evaluate the proposed method.The experiment demonstrates that 69%-92% of irrelevant sequences are pruned using the proposed method.

SINGULAR SPECTRUM ANALYSIS:METHODOLOGY AND APPLICATION TO ECONOMICS DATA

This paper describes the methodology of singular spectrum analysis (SSA) and demonstratethat it is a powerful method of time series analysis and forecasting,particulary for economic time series.The authors consider the application of SSA to the analysis and forecasting of the Iranian nationalaccounts data as provided by the Central Bank of the Islamic Republic of Iran.