研究时间序列异常模式挖掘具有重要的学术价值和现实意义。针对时间序列连续、非线性、高维的复杂结构，探索了时间序列异常模式挖掘的新途径。研究了基于分类的时间序列异常检测方法及特征属性降维方法，并应用于矿井瓦斯异常检测；同时提出了多传感器空间布局优化方法，开发了原型系统。主要研究内容如下： 针对逆分类方法无法用于连续属性的缺点，提出了一种连续属性逆分类的方法。它是一种同时对训练样本和测试样本进行分析的方法，其主要思想是：首先找到与类别属性相关的特征属性组合；然后将连续属性离散化并且为训练样本构造逆统计；最后对测试样本进行异常检测或遗失值估计。在IRIS和Ecoli数据集上的实验结果表明，该算法的分类准确率、平均相对误差和最小相对误差均优于朴素贝叶斯和ISGNN。 针对现有流形学习方法用于异常检测低准确率的缺点，提出了两种有监督的流形学习方法PSLLE和PQSLLE，通过计算偏向概率来调节样本间的距离，使得测试样本所隐含类别信息在其低维嵌入空间仍得以保存，其主要思想是：首先通过偏向概率计算近邻矩阵；然后计算权值矩阵；最后通过局部重建权值矩阵和近邻矩阵映射到低维嵌入空间。在IRIS、Wine和Semeion手写数字数据集上的实验结果表明，采用PSLLE和PQSLLE降维后进行KNN异常检测的准确率优于LLE和SLLE。 多传感器时间序列的异常模式是通过分析传感器的量测值得到的，表现为数据相关，缺少物理上的相关分析。针对这个不足，提出了多传感器的空间布局的优化方法，给出了形式化描述，并分别建立了多传感器在一维、二维和三维空间的布局优化模型。 将时间序列异常模式挖掘方法用于煤矿瓦斯异常检测中，开发了煤矿瓦斯涌出量异常检测系统，包括异常检测、遗失值估计和流形学习模块。测试结果表明系统运行正确。

Study of key technologies on abnormal pattern mining in time series has important academic value and practical significance. In order to deal with the continuity, non-linear and high-dimension of time series, this thesis aims to explore a new way to mine abnormal pattern of times series. The methods of abnormal detection based on classfication and dimensionality reduction of characters attributes are studied, which are applied to anomaly detection of coal mine gas datas. At the same time, the solutions and definition of layout of multisensor system are proposed. An abnormal detection software system of gas based on time series data mining is developed. The main contents are as follows: An inverse classification method of quantitative attributes was presented, which overcomes the disadvantage of most inverse classification algorithms address discrete attributes. The algorithm puts emphasis on analysis of the training samples and the test samples, the main idea is: firstly, a group of feature attributes are selected by using feature selection algorithm; then, the quantitative attributes are discretized by using discretization algorithms, and the inverted statistics are constructed on the training samples; finally, the test samples are analyzed by using the inverted statistics, and the method is applied to abnormal detection and estimating the missing values. Experimental results on IRIS and Ecoli datasets show that the accuracy of classification, the average relative deviation and the maximum relative deviation are better than KNN(k-Nearest Neighbors) and ISGNN(Iteration Self-Generated Neural Network). In order to overcome the low accuracy of anomaly detection based on manifold learning methods. Two supervised manifold learning algorithms are proposed, named PSLLE (Supervised Locally Linear Embedding based on Problitity) and PQSLLE (Quick Supervised Locally Linear Embedding based on Problitity). They both adjust the distance between samples by calculating the bias problitity of samples in order to preserve the hidden classlabels. So the test samples could be deal with by supervised manifold learning algorithm. Experimental results on IRIS, Wine and Semeion Handwritten Digit datasets show that the accuracy of anomaly detection after dimensionality reduction by proposed algorithm is better than LLE(Locally Linear Embedding) and SLLE(Supervised Locally Linear Embedding). Anomaly pattern multisensor time series is got by analysis of sensor values, which shows correlation of datas, and lacks physical correlation analysis. In order to overcome the disadvantage, the solutions and definition of layout of multisensor system are proposed. Formal description of the solutions is given, and the multi-objective optimization models of multi-sensor separately in one, tow, three dimensional spaces are established. Solve the problem of anomaly detection about coal mine gas by using methods of abnormal pattern mining in time series. Based on above research results, an abnormal detection system for gas data is developed, which includes three modules, abnormal detection, estimating the missing values and manifold learning. The test results show that the software runs well.