互信息与模糊C均值聚类集成的特征优选方法

针对大型数据中大量冗余特征的存在可能降低数据分类性能的问题,提出了一种基于互信息(MI)与模糊C均值(FCM)聚类集成的特征自动优选方法FCC-MI。首先分析了互信息特征及其相关度函数,根据相关度对特征进行排序;然后按照最大相关度对应的特征对数据进行分组,采用FCM聚类方法自动确定最优特征数目;最后基于相关度对特征进行了优选。在UCI机器学习数据库的7个数据集上进行实验,并与相关文献中提出的基于类内方差与相关度结合的特征选择方法(WCMFS)、基于近似Markov blanket和动态互信息的特征选择算法(B-AMBDMI)及基于互信息和遗传算法的两阶段特征选择方法(T-MI-GA)进行对比。理论分析和实验结果表明,FCC-MI不但提高了数据分类的效率,而且在有效保证分类精度的同时能自动确定最优特征子集,减少了数据集的特征数目,适用于海量、数据特征相关性大的特征约简及数据分析。     

基于模板的数控冲床加工路径排序方法研究

给出了满足实际加工要求的各种加工路径排序方式,提出了排序总体设计原理和流程.着重阐述了将二元关键码转换成一元关键码实现按扫描线排序、用遗传算法实现最短路径排序、按刀具面积大小实现综合排序的方法及设计过程.最后,通过开发应用实例验证了这种在数控床CAD/CAM软件设计系统中基于模板的加工路径排序设计方法是科学有效的.     

基于文本与视觉特征聚合的图像排序学习模型

图像搜索中重要的问题之一是如何有效地对搜索结果进行排序.现有图像搜索引擎的排序模型一般都基于相关文本而没有考虑图像的视觉特征.由于文本特征有时并不能很好地匹配图像的内容,所以搜索结果中会包含被错误排序的图像.针对该问题已经提出了视觉重排序方法,通过视觉信息来精炼基于文本的搜索结果.然而视觉重排序带来的性能提升有限,主要原因是基于文本的搜索结果中的错误会传播到视觉重排序阶段.本文基于排序学习的框架提出一个联合文本和视觉特征的图像排序学习模型,同时考虑了视觉和文本特征来进行排序学习,避免了视觉重排序中的错误传播.实验结果表明本文提出的排序模型显著地好于现有的重排序方法.     

一种基于判别式重排序的拼写校正方法

提出一种基于判别模型的拼写校正方法.它针对已有拼写校正系统Aspell的输出进行重排序,使用判别模型Ranking SVM来改进其性能.将现今较为成熟的拼写校正技术(包括编辑距离、基于字母的n元语法、发音相似度和噪音信道模型)以特征的形式整合到该模型中来,显著地提高了基准系统Aspell的初始排序质量,同时性能也超过了一些商用系统(如Microsoft Word 2003)的拼写校正模块.此外,还提出了一种在搜索引擎查询日志链中自动抽取拼写校正训练对的方法.基于这种方法训练的模型获得了基于人工标注数据所得结果相近的性能,它们分别将基准系统的错误率降低了32.2%和32.6%.     

基于RF的排序学习方法在电影数据集中的应用

针对自制电影数据集中电影的排序问题,文章提出了一种基于RF的Bootstrap自适应双集成排序学习方法(RandomForest-based Bootstrap Self-adaptive Double-ensemble,RF-based BSD).先利用电影媒体网站数据构建21个特征自建基于排序学习格式的电影数据集,BSD会根据输入数据集的查询数、查询-电影对数和特征数,通过Bootstrap自适应函数自动确定RF的子采样比例,然后使用单集成模型(比如MART,Multiple Additive Regression Tree,多重累计回归树)作为基学习器进行训练,最后采用bagging思想输出最终的双集成模型.实验结果显示,对比两个评价指标NDCG(Normalized Discounted Cumulative Gain,归一化折扣累计增益)和MAP(Mean Average Precision,平均值均值)的评估效果,发现BSD输出的双集成模型比单集成模型在两项指标上均有1％-3％左右的提升.     

基于特征选择和点互信息剪枝的产品属性提取方法

产品属性的自动抽取是情感分析中的重要研究内容.文中提出一种基于特征选择和词频及点互信息剪枝的产品属性提取方法.首先引入在分类任务中常用的l1-norm正则化(Lasso)方法,将产品属性抽取问题转换为分类中的特征选择问题,利用Lasso生成稀疏模型的特性,将模型中少量的特征作为产品特征属性候选集.然后根据候选特征属性集中的特征属性在文本中出现的频率进行排序并剪枝.最后经过进一步合并和点互信息剪枝处理,得到最终的产品属性集.在中文产品评论集上的实验证实文中方法的有效性.     

基于运动想象脑电信号的导联排序研究

为在脑机接口系统BCI(brain-computer interface)中有效选择导联进行特征提取和分类提供依据,研究了基于运动想象脑电信号的导联排序.根据公共空间模式算法CSP(common spatial pattern)原理提出了一种导联排序方法--基于协方差和主成分分析的排序算法CPSorting(covariance and principal component sorting),并研究了运动想象脑电信号MI(motor imagery)导联的排序情况以及排序靠前的导联对分类的贡献.利用公共空间模式算法对CPSorting排序后导联的数据提取特征,再分别应用支持向量机SVM和K近邻算法KNN进行分类.实验结果表明了该排序算法能有效地对基于运动想象脑电信号的导联进行排序.     

双重驱动的果蝇优化算法及其在PID控制器中的应用

针对工业环境中齿轮箱多故障特征难以选择的问题,结合Fisher Score与最大信息系数(MIC)构建一种新的故障特征优化选择方法.首先,考虑到多故障特征分布不均匀和重叠性问题,采用Fisher Score计算方法构建特征指标重要度排序规则;其次,在考虑冗余特征对有效特征表征的影响基础上,利用最大信息系数构建特征间关联性评价方法,对冗余特征实现更新排序;再次,以分类准确率为判断依据,基于支持向量机理论(SVM)对排序模型进行修正,建立基于Fisher Score与最大信息系数的故障特征优化选择方法;最后,利用UCI标准数据集和实验仿真的齿轮箱故障数据进行实验以验证所提出算法的有效性和工程实用性.仿真实验对比分析表明,与传统的mRMR、reliefF方法相比,所提出的方法特征子集数量适中,准确率更高.     

基于矩阵分解优化的排序学习特征构造方法

在排序学习中引入特征选择可以提高学习的效率和准确率。出于对选择速度的考虑,当前的研究主要从特征选择的角度出发,根据特征对排序的作用和特征之间的相似性选择对排序区分度最大的特征集合。由于特征大都是人工归纳的,因此特征和特征之间难免存在重叠和冗余。为了减少特征之间的冗余,从特征生成的角度出发,对现有特征进行矩阵分解,从而生成新的特征集。考虑到使用奇异值分解(Singular Value Decomposition SVD)等方法进行矩阵分解时不能综合考虑排序结果对特征的影响,基于特征矩阵对排序的效果、特征矩阵与原矩阵之间的差距来构造优化算法,提出了一种基于矩阵分解的排序学习优化方法,并根据该优化方法设计了排序学习特征选择算法MFRank。实验中使用映射随机梯度下降法近似求得优化问题的最优值,在公开测试集MQ2008上的结果显示,所提MFRank方法获得了与当前最优的特征选择方法即RankBoost和RankSVM-Struct等排序算法相当的结果。     

一种使用多Filter初始化GA种群的混合特征选择模型

特征选择已经是高维数据处理尤其是模式识别领域中的一个关键问题.提出一种混合特征选择模型用于从潜在的相关特征中选择那些最重要的特征.该模型包括两部分:filter部分与wrapper部分.在filter部分,4种不同的Filter方法分别对候选特征进行独立排序,在融合后进一步生成综合特征排序,综合排序随后产生遗传算法(GA)的初始种群.在wrapper部分,GA算法根据神经网络的分类准确率对个体(特征子集)进行评价,以便于搜索到最优的特征子集.测试结果表明,该模型不仅能有效地减少特征子集的大小,而且还可以进一步提高分类识别的准确率和效果.     

用于柔性生产线方案设计的分层工艺规划方法

在建立面向对象工艺模型的基础上，提出用于柔性生产线方案设计的分层工艺规划方法，即特征工艺设计、零件装夹定位规划、特征排序规划、工步排序规划、工序聚合规划等5个层次．采用基于黑板的递阶式层次推理结构，实现了综合排序策略，完成了用于柔性生产线方案设计的CAPP系统开发。     

A constraint-based operation sequencing for a knowledge-based process planning

Process planning is a decision-making process. Decisions on machining operations for a particular feature have to be made on various independent conditions such as which operation should be performed with which tools and under what cutting parameters. An integrated knowledge-based CAPP system called ProPlanner has been developed. The system has five modules namely information acquisition, feature recognition, machining operation planning and tool selection, set-up planning, and operation sequencing. Most process-planning systems do not produce alternative process plans. Usually, a fixed sequence created by a process plan is not necessarily the best possible sequence. Therefore, the aim should be to generate all possible operation sequences and use some optimality criteria to obtain the best sequence for the given operating environment. This paper presents an efficient heuristic algorithm, belongs to the system's operation sequencing module, for finding near-optimal operation sequences from all available process plans in a machining set-up. The costs of the various machining schemes are calculated and the machining scheme with the lowest cost is chosen. All feasible cutting tools are identified for each particular feature and the corresponding machining operations. This process is repeated for all the features in the machining set-up. All possible feature sequence combinations allowed by the current feature constraints are then generated. Appropriate cutting tools are identified and assigned to different operations. The feature sequence with the smallest number of tool changes is adopted.     

Exact Recognition of Compound Features by Feature Adjacency Matrix Elimination Algorithm

Aiming at the axiom of design for manufacture (DFM), this paper describes a recognition method for abstracting compound features from a part model and discloses the basic mechanism of compounding, also builds the corresponding 2D-simulation model. The inner association between feature neighboring and feature compounding is deeply discussed and, based on the essential transforming rule of two neighboring features, the corresponding feature adjacency matrix (FAM) of multi - feature entities are generated. For the manufacturing feature converted from the pure design feature; an innovative concept-homogenous compounding is presented to clarify the architecture of machining domain. Then, the FAM recurrence elimination algorithm is developed to determine all the compound features, and according to machining sequence, outputs a group of machining domains.     

Feature Setup Determination in Integrated CAD/CAM System For Concurrent Engineering

１ＩｎｔｒｏｄｕｃｔｉｏｎＷｉｔｈｔｈｅｇｒｏｗｉｎｇｃｏｍｐｅｔｉｔｉｏｎｉｎｔｈｅｍａｎｕｆａｃｔｕｒｅｉｎｄｕｓｔｒｙ，ｔｈｅｄｅｖｅｌｏｐｉｎｇｐｅｒｉｏｄｏｆｐｒｏｄｕｃｔｉｓｓｈｏｒｔｅｒａｎｄｓｈｏｒｔｅｒ．Ｓｏａｍａｎｕｆａｃｔｕｒｉ...     

五轴数控31/21/2轴倾斜面加工算法

针对五轴机床对带倾斜面类零件加工时编程难度大且精度较低的问题,采用31/21/2轴加工方法,通过定义可以简化加工编程的空间特性坐标系建立了任意结构五轴机床31/21/2轴加工的运动学模型.利用齐次变换矩阵和正向、逆向运动学相结合的方法详细论述了基于特性坐标系的适合三种典型五轴机床的31/21/2轴加工算法.此算法把31/21/2轴加工的空间编程问题转化为倾斜面平面编程问题,不仅简化编程而且可提高机床加工能力和生产效率.通过仿真实验验证了该算法的可靠性和高效性.     

CAPP中工艺规程主干生成的约束匹配算法

工艺决策是CAPP系统开发的关键，文中提出了一种新的工艺规程主干生成的约束匹配算法，该算法将工艺决策任务分配为特征加工链生成、工序排序和工步排序三个相对独立的子任务，每个子任务只解决一个主要问题而与其它因素无关，并分别采取了不同的子任务求解策略，有效地缩小了排序规模，极大地降低了算法复杂度，该算法的关键技术在RJ-CAPP系统中得到实现。     

Using Hopfield neural networks for operational sequencing for prismatic parts on NC machines

A numerical control (NC) machine is accurate and expensive equipment that provides us with flexible and reliable operations. However, many process planners only use their instinct in planning operational sequencing and do not minimize non-cutting time. In this paper, the sequencing task is formulated as constrained optimization problems to generate efficient machining of a part for NC machines. Factors in this study include part and table orientations and feature grouping for same cutting tools. First, this proposed method finds the minimum part orientations to cover all part features in order to reduce the most time consuming setups. Then it finds the minimum table orientations needed based on the accessibility of parts features in each part orientation. Most importantly, the preliminary sequence is refined by including feature precedence relationship and feature clustering for tools and tool approaching directions that will reduce tool re-orientation and tool changing time. Due to potential conflicts of constraints for sequencing optimization from the imbedding of precedence relationships, the soft computing ability of neural networks must be utilized in this refining procedure. This paper models the problem that allows an analogy to be conducted between finding the best operation sequence and minimizing the energy function of a Hopfield neural network. Finally, a spindle cover is used as an example to illustrate the implementation of the proposed method.     

机械加工MBD毛坯模型的特征识别设计方法

为了解决基于模型的定义(MBD)环境下机械加工毛坯辅助设计问题,采用特征识别技术辅助创建三维毛坯模型.首先分析了机械加工工艺设计环节MBD毛坯模型的组成要素与模型要求,并建立零件模型的属性面邻接图;根据零件模型上加工特征的特点划分简单特征、体积特征和表面特征;通过属性面邻接子图的模式从特征集合中识别出简单特征与体积特征,并利用补特征法与半空间法抑制这2种特征;最后,根据工艺员输入的表面加工余量创建抑制表面特征的实体.在整个毛坯设计推理过程中研究了三维制造标注的维护方法,最终达到辅助工艺员快速设计MBD毛坯模型的目的.     

Machining process sequencing with fuzzy expert system and genetic algorithms

Traditional process planning systems are usually established in a deterministic framework that can only deal with precise information. However, in a practical manufacturing environment, decision making frequently involves uncertain and imprecise information. This paper describes a fuzzy approach for solving the process selection and sequencing problem under uncertainty. The proposed approach comprises a two-stage process for machining process selection and sequencing. The two stages are called intra-feature planning and inter-feature planning, respectively. According to the feature precedence relationship of a machined part, the intra-feature planning module generates a local optimal operation sequence for each feature element. This is based on a fuzzy expert system incorporated with genetic algorithms for machining cost optimization according to the cost-tolerance relationship. Manufacturing resources such as machines, tools, and fixtures are allocated to each selected operation to form an Operation-Machine-Tool (OMT) unit in the manufacturing resources allocation module. Finally, inter-feature planning generates a global OMT sequence. A genetic algorithm with fuzzy numbers and fuzzy arithmetic is developed to solve this global sequencing problem.