复杂系统故障早期多变量序列的混沌模式分析和预测

项目名称： 复杂系统故障早期多变量序列的混沌模式分析和预测

项目编号： No.60804025

项目类型： 青年科学基金项目

立项/批准年度： 2009

项目学科： 数理科学和化学

项目作者： 席剑辉

作者单位： 沈阳航空航天大学

项目金额： 19万元

中文摘要： 系统某些故障在演化初期很难被及时识别。课题基于混沌理论，研究异常信息检测方法，为制定科学的维修机制打下基础。主要有：(1)异常信息的混沌特征描述和提取。从去噪角度提出一种新的小波半软阈值函数，在阈值范围内对小波系数进行细化处理，改进去噪效果。另一方面，针对含噪序列，引入对高斯噪声鲁棒性强的高阶累积量计算特性参数，提高抗噪能力。(2)多元变量相空间重构。首先应用单变量相空间重构思想，建立一个包含多个变量信息的初始向量；采用FastICA方法提取独立元，进行降维。提出一种基于高阶累积量的局部本征维数计算方法，针对含噪序列重构相空间。(3)多模型组合实现故障早期预测。对径向基函数(RBF)网络引入线性相关函数和非线性相关函数，确定能够包含系统线性、非线性有效信息的小数据集，优化网络训练样本；结合遗传算法全局搜索最优的聚类中心宽度系数，优化网络训练过程；利用RBF网络对BP网络和支持向量机进行加权组合，提高模型泛化能力。对最小二乘支持向量机(LSSVM)提出混沌粒子群优化算法和交叉验证算法相结合的参数寻优方法，提高LSSVM的学习和泛化能力。已发表和录用论文20余篇，含EI检索9篇。

中文关键词： 故障模式；混沌特性；多变量序列；预测

英文摘要： It is very difficult for some failures in the system to be identified at their early stage. Based on chaotic theory, this project studies detection method for abnormal information, and lay a foundation for making scientific maintenance mechanism. The main works include: (1) Description and extraction of chaotic characteristics from abnormal information. In view of de-noising, a new wavelet half-soft threshold function is proposed to refine the wavelet coefficients within the threshold, which improves the de-noising results. On the other hand, for noisy series, higher-order cumulant, which robustness is strong for Gaussian white noise, is introduced to compute characteristic parameters and improve the anti-noise ability. (2) Multivariable phase space reconstruction. First apply single variable phase space reconstruction thought to form an initial vector including multivariable information. Then use FastICA method to extract independent components, and realize dimension reduction. An algorithm for calculating local intrinsic dimension (LID) based on high-order cumulant is proposed to reconstruct phase space for noisy series. (3) Early precdiction of faults using multi-model combination. For the radial basis function (RBF) network, a linear function and a nonlinear function are introduced to define a small data set which includes effective linear and nonlinear information of the system. The training samples of network are optimized. The training process is also optimized by combination with the genetic algorithm to globally search the optimal width parameter of clustering centers. Weighted combination of BP network and support vector machine using the RBF network can improve generalization ablility of the model. For the least square support vector machine (LSSVM), chaos particle swarm optimization algorithm combined with k-fold cross validation is proposed for selecting the optimal parameters, therefore improve the learning and generalization ability. Studies have been published and accepted more than 20 papers, including 9 papers indexed by EI.

英文关键词： failure mode; chaotic characteristics; multivariable series; prediction