王惮维-Control Conference (CGNCC2014)
机器人运动控制方法综述
机器人运动控制方法综述摘要：随着工业的不断发展，机器人的应用领域日益广泛，如汽车、飞机、发电机等零部件的焊接、磨抛、装配等。
在这些任务中，机器人的运动精度对提高产品质量至关重要。
然而，机器人强耦合、非线性、性能依赖位形，且实际运动过程中存在摩擦、扰动等多种不确性复杂因素，因此对机器人运动控制方法进行分析整理具有重要意义和应用价值。
本文对现有主流机器人（主要针对机械臂）运动控制算法进行了综述，分析了这些算法的特点以及存在的不足，最后给出了机器人运动控制研究展望。
该工作可以为机器人运动控制方法的选择提供依据，具有一定的实用价值。
关键词：机器人，运动控制，综述1引言随着多传感和人工智能等技术的进步，机器人正日益广泛应用于焊接、磨抛、装配等智能制造领域，显著提升了生产效率，降低了生产成本。
因此，机器人是我国制造技术领域的主攻方向之一[1]，也是助力我国从制造大国向制造强国快速迈进的战略高技术之一。
当前，高精度运动是机器人领域发展的主要趋势之一，对形成高质量的操作品质至关重要，而高精度运动的实现要求控制方法具有高动态响应和强鲁棒性。
然而，机器人齿轮和连杆结构受载后均存在不同程度的变形，且整个系统本身强耦合、非线性、动态特性时变、操作工况不确定性复杂多样，导致高动态响应和强鲁棒性的运动控制方法设计尤为困难，从而产生较大控制偏差，影响机器人运动精度，造成操作质量下降。
基于上述分析，本文对机器人（主要针对机械臂）的主流运动控制方法进行了综述，分析了这些方法中存在的不足，并给出了机器人运动控制方法未来发展趋势的一些思考。
2机器人运动控制方法作为一种复杂自动化系统，机器人具有非线性、强耦合、多变量时变特性。
高速运动时各关节惯量变化较大，耦合强烈，低速时摩擦、饱和等非线性效应明显，这些极大地增加了控制难度。
为了使机器人系统稳定运行，要求控制系统能实时提供与机器人动力学特征和多源扰动特征相匹配的控制特性，这给予了运动控制方法极大的挑战。
Feature selection for high-dimensional data A fast correlation-based filter solution
Feature Selection for High-Dimensional Data:A Fast Correlation-Based Filter SolutionLei Yu leiyu@ Huan Liu hliu@ Department of Computer Science&Engineering,Arizona State University,Tempe,AZ85287-5406,USAAbstractFeature selection,as a preprocessing step tomachine learning,has been effective in reduc-ing dimensionality,removing irrelevant data,increasing learning accuracy,and improvingcomprehensibility.However,the recent in-crease of dimensionality of data poses a se-vere challenge to many existing feature se-lection methods with respect to efficiencyand effectiveness.In this work,we intro-duce a novel concept,predominant correla-tion,and propose a fastfilter method whichcan identify relevant features as well as re-dundancy among relevant features withoutpairwise correlation analysis.The efficiencyand effectiveness of our method is demon-strated through extensive comparisons withother methods using real-world data of highdimensionality.1.IntroductionFeature selection is frequently used as a preprocessing step to machine learning.It is a process of choosing a subset of original features so that the feature space is optimally reduced according to a certain evaluation criterion.Feature selection has been a fertilefield of research and development since1970’s and shown very effective in removing irrelevant and redundant fea-tures,increasing efficiency in learning tasks,improv-ing learning performance like predictive accuracy,and enhancing comprehensibility of learned results(Blum &Langley,1997;Dash&Liu,1997;Kohavi&John, 1997).In recent years,data has become increasingly larger in both rows(i.e.,number of instances)and columns(i.e.,number of features)in many applica-tions such as genome projects(Xing et al.,2001),text categorization(Yang&Pederson,1997),image re-trieval(Rui et al.,1999),and customer relationship management(Ng&Liu,2000).This enormity may cause serious problems to many machine learning al-gorithms with respect to scalability and learning per-formance.For example,high dimensional data(i.e., data sets with hundreds or thousands of features),can contain high degree of irrelevant and redundant infor-mation which may greatly degrade the performance of learning algorithms.Therefore,feature selection be-comes very necessary for machine learning tasks when facing high dimensional data nowadays.However,this trend of enormity on both size and dimensionality also poses severe challenges to feature selection algorithms. Some of the recent research efforts in feature selection have been focused on these challenges from handling a huge number of instances(Liu et al.,2002b)to deal-ing with high dimensional data(Das,2001;Xing et al., 2001).This work is concerned about feature selection for high dimensional data.In the following,wefirst review models of feature selection and explain why a filter solution is suitable for high dimensional data,and then review some recent efforts in feature selection for high dimensional data.Feature selection algorithms can broadly fall into the filter model or the wrapper model(Das,2001;Kohavi &John,1997).Thefilter model relies on general char-acteristics of the training data to select some features without involving any learning algorithm,therefore it does not inherit any bias of a learning algorithm.The wrapper model requires one predetermined learning al-gorithm in feature selection and uses its performance to evaluate and determine which features are selected. As for each new subset of features,the wrapper model needs to learn a hypothesis(or a classifier).It tends to give superior performance as itfinds features better suited to the predetermined learning algorithm,but it also tends to be more computationally expensive(Lan-gley,1994).When the number of features becomes very large,thefilter model is usually a choice due to its computational efficiency.Proceedings of the Twentieth International Conference on Machine Learning(ICML-2003),Washington DC,2003.To combine the advantages of both models,algorithms in a hybrid model have recently been proposed to deal with high dimensional data(Das,2001;Ng,1998;Xing et al.,2001).In these algorithms,first,a goodness measure of feature subsets based on data characteris-tics is used to choose best subsets for a given cardinal-ity,and then,cross validation is exploited to decide a final best subset across different cardinalities.These algorithms mainly focus on combiningfilter and wrap-per algorithms to achieve best possible performance with a particular learning algorithm at the same time complexity offilter algorithms.In this work,we focus on thefilter model and aim to develop a new feature selection algorithm which can effectively remove both irrelevant and redundant features and is less costly in computation than the current available algorithms.In section2,we review current algorithms within the filter model and point out their problems in the context of high dimensionality.In section3,we describe corre-lation measures which form the base of our method in evaluating feature relevance and redundancy.In sec-tion4,wefirst propose our method which selects good features for classification based on a novel concept, predominant correlation,and then present a fast algorithm with less than quadratic time complexity.In section5,we evaluate the efficiency and effectiveness of this algorithm via extensive experiments on various real-world data sets comparing with other representa-tive feature selection algorithms,and discuss the im-plications of thefindings.In section6,we conclude our work with some possible extensions.2.Related WorkWithin thefilter model,different feature selection al-gorithms can be further categorized into two groups, namely feature weighting algorithms and subset search algorithms,based on whether they evaluate the good-ness of features individually or through feature sub-sets.Below,we discuss the advantages and shortcom-ings of representative algorithms in each group. Feature weighting algorithms assign weights to fea-tures individually and rank them based on their rel-evance to the target concept.There are a number of different definitions on feature relevance in machine learning literature(Blum&Langley,1997;Kohavi& John,1997).A feature is good and thus will be se-lected if its weight of relevance is greater than a thresh-old value.A well known algorithm that relies on rele-vance evaluation is Relief(Kira&Rendell,1992).The key idea of Relief is to estimate the relevance of fea-tures according to how well their values distinguish be-tween the instances of the same and different classes that are near each other.Relief randomly samples a number(m)of instances from the training set and up-dates the relevance estimation of each feature based on the difference between the selected instance and the two nearest instances of the same and opposite classes.Time complexity of Relief for a data set with M instances and N features is O(mMN).With m be-ing a constant,the time complexity becomes O(MN), which makes it very scalable to data sets with both a huge number of instances and a very high dimension-ality.However,Relief does not help with removing redundant features.As long as features are deemed relevant to the class concept,they will all be selected even though many of them are highly correlated to each other(Kira&Rendell,1992).Many other algo-rithms in this group have similar problems as Relief does.They can only capture the relevance of features to the target concept,but cannot discover redundancy among features.However,empirical evidence from fea-ture selection literature shows that,along with irrele-vant features,redundant features also affect the speed and accuracy of learning algorithms and thus should be eliminated as well(Hall,2000;Kohavi&John,1997). Therefore,in the context of feature selection for high dimensional data where there may exist many redun-dant features,pure relevance-based feature weighting algorithms do not meet the need of feature selection very well.Subset search algorithms search through candidate feature subsets guided by a certain evaluation mea-sure(Liu&Motoda,1998)which captures the good-ness of each subset.An optimal(or near optimal)sub-set is selected when the search stops.Some existing evaluation measures that have been shown effective in removing both irrelevant and redundant features in-clude the consistency measure(Dash et al.,2000)and the correlation measure(Hall,1999;Hall,2000).Con-sistency measure attempts tofind a minimum num-ber of features that separate classes as consistently as the full set of features can.An inconsistency is de-fined as two instances having the same feature values but different class labels.In Dash et al.(2000),dif-ferent search strategies,namely,exhaustive,heuristic, and random search,are combined with this evalua-tion measure to form different algorithms.The time complexity is exponential in terms of data dimension-ality for exhaustive search and quadratic for heuristic search.The complexity can be linear to the number of iterations in a random search,but experiments show that in order tofind best feature subset,the number of iterations required is mostly at least quadratic to the number of features(Dash et al.,2000).In Hall(2000), a correlation measure is applied to evaluate the good-ness of feature subsets based on the hypothesis that a good feature subset is one that contains features highly correlated with the class,yet uncorrelated with each other.The underlying algorithm,named CFS,also exploits heuristic search.Therefore,with quadratic or higher time complexity in terms of dimensionality, existing subset search algorithms do not have strong scalability to deal with high dimensional data.To overcome the problems of algorithms in both groups and meet the demand for feature selection for high dimensional data,we develop a novel algorithm which can effectively identify both irrelevant and redundant features with less time complexity than subset search algorithms.3.Correlation-Based MeasuresIn this section,we discuss how to evaluate the good-ness of features for classification.In general,a feature is good if it is relevant to the class concept but is not redundant to any of the other relevant features.If we adopt the correlation between two variables as a good-ness measure,the above definition becomes that a fea-ture is good if it is highly correlated with the class but not highly correlated with any of the other features. In other words,if the correlation between a feature and the class is high enough to make it relevant to(or predictive of)the class and the correlation between it and any other relevant features does not reach a level so that it can be predicted by any of the other relevant features,it will be regarded as a good feature for the classification task.In this sense,the problem of fea-ture selection boils down tofind a suitable measure of correlations between features and a sound procedure to select features based on this measure.There exist broadly two approaches to measure the correlation between two random variables.One is based on classical linear correlation and the other is based on information theory.Under thefirst approach, the most well known measure is linear correlation co-efficient.For a pair of variables(X,Y),the linear correlation coefficient r is given by the formular=i(x i−x i)(y i−y i)i(x i−x i)2i(y i−y i)2(1)where x i is the mean of X,and y i is the mean of Y. The value of r lies between-1and1,inclusive.If X and Y are completely correlated,r takes the value of 1or-1;if X and Y are totally independent,r is zero. It is a symmetrical measure for two variables.Other measures in this category are basically variations of the above formula,such as least square regression er-ror and maximal information compression index(Mi-tra et al.,2002).There are several benefits of choos-ing linear correlation as a feature goodness measure for classification.First,it helps remove features with near zero linear correlation to the class.Second,it helps to reduce redundancy among selected features. It is known that if data is linearly separable in the original representation,it is still linearly separable if all but one of a group of linearly dependent features are removed(Das,1971).However,it is not safe to always assume linear correlation between features in the real world.Linear correlation measures may not be able to capture correlations that are not linear in nature. Another limitation is that the calculation requires all features contain numerical values.To overcome these shortcomings,in our solution we adopt the other approach and choose a correlation measure based on the information-theoretical concept of entropy,a measure of the uncertainty of a random variable.The entropy of a variable X is defined asH(X)=−iP(x i)log2(P(x i)),(2)and the entropy of X after observing values of another variable Y is defined asH(X|Y)=−jP(y j)iP(x i|y j)log2(P(x i|y j)),(3) where P(x i)is the prior probabilities for all values of X,and P(x i|y i)is the posterior probabilities of X given the values of Y.The amount by which the en-tropy of X decreases reflects additional information about X provided by Y and is called information gain(Quinlan,1993),given byIG(X|Y)=H(X)−H(X|Y).(4)According to this measure,a feature Y is regarded more correlated to feature X than to feature Z,if IG(X|Y)>IG(Z|Y).About information gain mea-sure,we have the following theorem.Theorem Information gain is symmetrical for two random variables X and Y.Proof Sketch:To prove IG(X|Y)=IG(Y|X),we need to prove H(X)−H(X|Y)=H(Y)−H(Y|X). This can be easily derived from H(X,Y)=H(X)+ H(Y|X)=H(Y)+H(X|Y).Symmetry is a desired property for a measure of cor-relations between features.However,information gain is biased in favor of features with more values.Fur-thermore,the values have to be normalized to ensurethey are comparable and have the same affect.There-fore,we choose symmetrical uncertainty(Press et al., 1988),defined as follows.SU(X,Y)=2IG(X|Y)H(X)+H(Y)(5)It compensates for information gain’s bias toward fea-tures with more values and normalizes its values to the range[0,1]with the value1indicating that knowledge of the value of either one completely predicts the value of the other and the value0indicating that X and Y are independent.In addition,it still treats a pair of features symmetrically.These entropy-based mea-sures require nominal features,but they can be applied to measure correlations between continuous features as well,if the values are discretized properly in ad-vance(Fayyad&Irani,1993;Liu et al.,2002a).There-fore,we use symmetrical uncertainty in this work. 4.A Correlation-Based Filter Approach4.1.MethodologyUsing symmetrical uncertainty(SU)as the goodness measure,we are now ready to develop a procedure to select good features for classification based on corre-lation analysis of features(including the class).This involves two aspects:(1)how to decide whether a fea-ture is relevant to the class or not;and(2)how to decide whether such a relevant feature is redundant or not when considering it with other relevant features. The answer to thefirst question can be using a thresh-old SU value decided by the user,as the method used by many other feature weighting algorithms(e.g.,Re-lief).More specifically,suppose a data set S contains N features and a class C.Let SU i,c denote the SU value that measures the correlation between a feature f i and the class C(named c-correlation),then a subset S of relevant features can be decided by a threshold SU valueδ,such that∀f i∈S ,1≤i≤N,SU i,c≥δ. The answer to the second question is more complicated because it may involve analysis of pairwise correlations between all features(named f-correlation),which re-sults in a time complexity of O(N2)associated with the number of features N for most existing algorithms. To solve this problem,we propose our method below. Since f-correlations are also captured by SU values, in order to decide whether a relevant feature is redun-dant or not,we need tofind a reasonable way to decide the threshold level for f-correlations as well.In other words,we need to decide whether the level of correla-tion between two features in S is high enough to cause redundancy so that one of them may be removed from S .For a feature f i in S ,the value of SU i,c quantifies the extent to which f i is correlated to(or predictive of)the class C.If we examine the value of SU j,i for ∀f j∈S (j=i),we will also obtain quantified es-timations about the extent to which f i is correlated to(or predicted by)the rest relevant features in S . Therefore,it is possible to identify highly correlated features to f i in the same straightforward manner as we decide S ,using a threshold SU value equal or sim-ilar toδ.We can do this for all features in S .How-ever,this method only sounds reasonable when we try to determine highly correlated features to one concept while not considering another concept.In the con-text of a set of relevant features S already identified for the class concept,when we try to determine the highly correlated features for a given feature f i within S ,it is more reasonable to use the c-correlation level between f i and the class concept,SU i,c,as a refer-ence.The reason lies on the common phenomenon-a feature that is correlated with one concept(e.g.,the class)at a certain level may also be correlated with some other concepts(features)at the same or an even higher level.Therefore,even the correlation between this feature and the class concept is larger than some thresholdδand thereof making this feature relevant to the class concept,this correlation is by no means pre-dominant.To be more precise,we define the concept of predominant correlation as follows.Definition1(Predominant correlation).The corre-lation between a feature f i(f i∈S)and the class C is predominant iffSU i,c≥δ,and∀f j∈S (j=i),there exists no f j such that SU j,i≥SU i,c.If there exists such f j to a feature f i,we call it a redundant peer to f i and use S Pito denote the set of all redundant peers for f i.Given f i∈S and S Pi (S Pi=∅),we divide S Piinto two parts,S Pi+andS Pi−,where SP i+={fj|f j∈S Pi,SU j,c>SU i,c}and S Pi−={fj|f j∈S Pi,SU j,c≤SU i,c}.Definition2(Predominant feature).A feature is pre-dominant to the class,iffits correlation to the class is predominant or can become predominant after remov-ing its redundant peers.According to the above definitions,a feature is good if it is predominant in predicting the class concept, and feature selection for classification is a process that identifies all predominant features to the class concept and removes the rest.We now propose three heuristics that together can effectively identify predominant fea-tures and remove redundant ones among all relevant features,without having to identify all the redundant peers for every feature in S ,and thus avoids pairwise analysis of f-correlations between all relevant features.Our assumption in developing these heuristics is that if two features are found to be redundant to each other and one of them needs to be removed,removing the one that is less relevant to the class concept keeps more information to predict the class while reducing redun-dancy in the data.Heuristic1(if S Pi +=∅).Treat fias a predom-inant feature,remove all features in S Pi −,and skipidentifying redundant peers for them.Heuristic2(if S Pi +=∅).Process all features inS Pi +before deciding whether or not to remove fi.Ifnone of them becomes predominant,follow Heuristic1; otherwise only remove f i and decide whether or not to remove features in S Pi−based on other features in S . Heuristic3(starting point).The feature with the largest SU i,c value is always a predominant feature and can be a starting point to remove other features.4.2.Algorithm and AnalysisBased on the methodology presented before,we de-velop an algorithm,named FCBF(Fast Correlation-Based Filter).As in Figure1,given a data set withinput:S(f1,f2,...,f N,C)//a training data set δ//a predefined threshold output:S best//an optimal subset1begin2for i=1to N do begin3calculate SU i,c for f i;4if(SU i,c≥δ)5append f i to Slist ;6end;7order Slist in descending SU i,c value;8f p=getF irstElement(Slist );9do begin10f q=getNextElement(Slist ,f p);11if(f q<>NULL)12do begin13f q=f q;14if(SU p,q≥SU q,c)15remove f q from Slist ;16f q=getNextElement(Slist ,f q);17else f q=getNextElement(Slist ,f q);18end until(f q==NULL);19f p=getNextElement(Slist ,f p);20end until(f p==NULL);21S best=Slist ;22end;Figure1.FCBF Algorithm N features and a class C,the algorithmfinds a setof predominant features S best for the class concept.It consists of two major parts.In thefirst part(line2-7), it calculates the SU value for each feature,selects rele-vant features into Slistbased on the predefined thresh-oldδ,and orders them in descending order according to their SU values.In the second part(line8-20),it further processes the ordered list Slistto remove redundant features and only keeps predominant ones among all the selected relevant features.According to Heuristic1,a feature f p that has already been deter-mined to be a predominant feature can always be used tofilter out other features that are ranked lower than f p and have f p as one of its redundant peers.The iteration starts from thefirst element(Heuristic3)inSlist(line8)and continues as follows.For all the re-maining features(from the one right next to f p to thelast one in Slist),if f p happens to be a redundant peerto a feature f q,f q will be removed from Slist(Heuris-tic2).After one round offiltering features based on f p,the algorithm will take the currently remaining fea-ture right next to f p as the new reference(line19)to repeat thefiltering process.The algorithm stops untilthere is no more feature to be removed from Slist. Thefirst part of the above algorithm has a linear time complexity in terms of the number of features N.As to the second part,in each iteration,using the pre-dominant feature f p identified in the previous round, FCBF can remove a large number of features that are redundant peers to f p in the current iteration.The best case could be that all of the remaining features following f p in the ranked list will be removed;the worst case could be none of them.On average,we can assume that half of the remaining features will be removed in each iteration.Therefore,the time com-plexity for the second part is O(N log N)in terms of N.Since the calculation of SU for a pair of features is linear in term of the number of instances M in a data set,the overall complexity of FCBF is O(MN log N).5.Empirical StudyThe objective of this section is to evaluate our pro-posed algorithm in terms of speed,number of selected features,and learning accuracy on selected features.5.1.Experiment SetupIn our experiments,we choose three representative fea-ture selection algorithms in comparison with FCBF. One is a feature weighting algorithm,ReliefF(an ex-tension to Relief)which searches for several nearest neighbors to be robust to noise and handles multi-ple classes(Kononenko,1994);the other two are sub-set search algorithms which exploit sequential forward search and utilizes correlation measure or consistency measure to guide the search,denoted as CorrSF and ConsSF respectively.CorrSF is a variation of the CFS algorithm mentioned in section2.The reason why we prefer CorrSF to CFS is because both ex-periments in Hall(1999)and our initial experiments show that CFS only produces slightly better results than CorrSF,but CorrSF based on sequential forward search runs faster than CFS based on bestfirst search with5nodes expansion and therefore is more suitable for high dimensional data.In addition to feature se-lection algorithms,we also select two different learning algorithms,C4.5(Quinlan,1993)and NBC(Witten& Frank,2000),to evaluate the accuracy on selected fea-tures for each feature selection algorithm.Table1.Summary of bench-mark data sets.Title Features Instances Classes Lung-cancer57323 Promoters591062 Splice6231903 USCensus906893383 CoIL20008658222 Chemical1519363 Musk216965982 Arrhythmia28045216 Isolet618156026 Multi-features650200010The experiments are conducted using Weka’s imple-mentation of all these algorithms and FCBF is also implemented in Weka environment(Witten&Frank, 2000).All together10data sets are selected from the UCI Machine Learning Repository(Blake&Merz, 1998)and UCI KDD Archive(Bay,1999).A summary of data sets is presented in Table1.For each data set,we run all four feature selection algorithms,FCBF,ReliefF,CorrSF,ConsSF,respec-tively,and record the running time and the number of selected features for each algorithm.We then ap-ply C4.5and NBC on the original data set as well as each newly obtained data set containing only the se-lected features from each algorithm and record overall accuracy by10-fold cross-validation.5.2.Results and DiscussionsTable2records the running time and the number of selected features for each feature selection algorithm. For ReliefF,the parameter k is set to5(neighbors)and m is set to30(instances)throughout the experiments.From Table2,we can observe that for each algorithm the running times over different data sets are consis-tent with our previous time complexity analysis.From the averaged values in the last row of Table2,it is clear that FCBF runs significantly faster(in degrees)than the other three algorithms,which verifies FCBF’s su-perior computational efficiency.What is interesting is that ReliefF is unexpectedly slow even though its time complexity becomes O(MN)with afixed sample size m.The reason lies on that searching for nearest neighbors involves distance calculation which is more time consuming than the calculation of symmetrical uncertainty value.From Table2,it is also clear that FCBF achieves the highest level of dimensionality reduction by selecting the least number of features(with only one exception in USCensus90),which is consistent with our theoret-ical analysis about FCBF’s ability to identify redun-dant features.Tables3and4show the learning accuracy of C4.5and NBC respectively on different feature sets.From the averaged accuracy over all data sets,we observe that, in general,(1)FCBF improves the accuracy of both C4.5and NBC;and(2)of the other three algorithms, only CorrSF can enhance the accuracy of C4.5to the same level as FCBF does.From individual accuracy values,we also observe that for most of the data sets, FCBF can maintain or even increase the accuracy. The above experimental results suggest that FCBF is practical for feature selection for classification of high dimensional data.It can efficiently achieve high degree of dimensionality reduction and enhance classification accuracy with predominant features.6.ConclusionsIn this paper,we propose a novel concept of predomi-nant correlation,introduce an efficient way of analyz-ing feature redundancy,and design a fast correlation-basedfilter approach.A new feature selection algo-rithm FCBF is implemented and evaluated through extensive experiments comparing with related feature selection algorithms.The feature selection results are further verified by applying two different classification algorithms to data with and without feature selec-tion.Our approach demonstrates its efficiency and effectiveness in dealing with high dimensional data for classification.Our further work will extend FCBF to work on data with higher dimensionality(thousands of features).We will study in more detail redun-dant features and their role in classification,and com-bine FCBF with feature discretization algorithms to smoothly handle data of different feature types.。
基于改进萤火虫算法的光伏最大功率点跟踪
工程数学模型:
I = IPV - I0
expæqV + qRSI - 1ö
è nkθ
ø
- V + RSI , RP
图 1 光伏电池等效电路 Fig. 1 Equivalent circuit of photovoltaic cell
其中:I 为光伏电池的输出电流;V 为光伏电池输出电压;IPV 为光生电流;I0 为二极管反向饱和电流;q 为电
光伏阵列封装模块与光伏阵列输出特性仿真模型见图 2、图 3. 在 θ 为 25 ℃ ,S 为 1 000 W / m2 的 STC 条件下测得光伏阵列 U-I 以及 P 曲线,见图 4、图 5. 由图 4、图 5 可见,光伏阵列存在一个最大功率点 (MPP) ,在不同环境下,使光伏阵列始终工作在最大功率点是 MPPT 技术要解决的关键问题. 采用好的控 制策略可以使光伏电池始终工作在最大功率点,提升系统经济性,提高太阳能利用率,最大限度地把太阳 能转换为光能.
式中:β0 为最大吸引度因子.
萤火虫空间位置更新:
xt+1 i
=
x
t i
+
βij(
r)
e
-
γr2
(
x
t j
-
xti )
+ α(rand
-
0. 5)
,
式中:α(rand - 0. 5) 代表一个随机扰动项,主要目的是扩大搜索空间. 2. 1. 2 萤火虫算法实现流程
萤火虫算法流程见图 6. 2. 2 改进萤火虫算法
xn+1 = cos( karccosxn) ,xn ∈ [ - 1,1] ,
式中:xn 与 xn+1 分别为当前与下一次向量值;k 为随机常数.
郑州大学硕士论文模板
学校代码 10459学号或申请号密级硕士学位论文论文题目作者姓名：张三导师姓名：李四教授学科门类：工科专业名称：培养院系：大电气、大物工完成时间：20xx年4月A thesissubmitted toZhengzhouUniversityfor the degree ofMasterThesis TitleBySan ZhangSupervisor: Prof. Si LiMajor NameInstitute NameApril20xx学位论文原创性声明本人郑重声明：所呈交的学位论文，是本人在导师的指导下，独立进行研究所取得的成果。
除文中已经注明引用的内容外，本论文不包含任何其他个人或集体已经发表或撰写过的科研成果。
对本文的研究作出重要贡献的个人和集体，均已在文中以明确方式标明。
本声明的法律责任由本人承担。
学位论文作者：日期：年月日学位论文使用授权声明本人在导师指导下完成的论文及相关的职务作品，知识产权归属郑州大学。
根据郑州大学有关保留、使用学位论文的规定，同意学校保留或向国家有关部门或机构送交论文的复印件和电子版，允许论文被查阅和借阅；本人授权郑州大学可以将本学位论文的全部或部分编入有关数据库进行检索，可以采用影印、缩印或者其他复制手段保存论文和汇编本学位论文。
本人离校后发表、使用学位论文或与该学位论文直接相关的学术论文或成果时，第一署名单位仍然为郑州大学。
保密论文在解密后应遵守此规定。
学位论文作者：日期：年月日摘要随着我国工业技术水平的快速发展与进步，对材料的质量的要求也越来越高。
在各个行业领域中，材料和器件中表面和亚表面的完整性会影响和决定系统和仪器设备的工作效率以及运行寿命。
此外，随着中国工业的高速发展，空气及水源污染问题也变得日益严峻。
环境污染导致了某些疾病的高发，皮肤病即是受环境因素诱导的典型疾病之一。
近年来，中国居民的皮肤病发病率逐年上升，且新增病例呈年轻化趋势。
因此实现对材料表面裂纹和皮肤表面病变的无损检测尤为重要。
APEC文艺演出多媒体视频设计综述
APEC文艺演出多媒体视频设计综述作者：王之纲来源：《演艺科技》2014年第12期【摘要】2014 APEC文艺演出中主创团队充分运用多媒体技术，以地面LED屏幕、高清彩幕、纱幕投影等多重影像媒介与山水画卷的舞美环境自然融合，构建了一个美轮美奂、变幻无穷的演出空间，体现出了当代中国的审美意趣和文化内涵。
【关键词】 APEC；多媒体；视频设计；高清彩幕；视频影像；创意文章编号： 10.3969/j.issn.1674-8239.2014.12.004【Abstract】The main team of the 2014 APEC art performance made full use of multimedia technology， with the natural fusion of multiple visual media like LED screen on the ground， high definition color curtain ， veil projection and the stage craft environment landscape picture scroll. A magnificent and constantly changing performance space was built， which reflects the aesthetic interest and cultural connotation of contemporary China.【Key Words】APEC； multimedia； video design； HD color screen； video images；originality近年来随着多媒体技术的迅速发展，国际范围内的重大活动，如奥运会的开闭幕式以及大型文艺演出中，多媒体已经成为一个不可或缺的重要设计元素。
考虑节能的拓扑聚合虚拟网映射算法
考虑节能的拓扑聚合虚拟网映射算法王博;陈庶樵;王志明;王文钊【摘要】网络虚拟化的关键问题是虚拟网映射,能耗开销的快速增长使得节能成为底层设施供应商关注的目标.针对虚拟网映射中的节能问题,提出一种集中使用网络拓扑的节能虚拟网映射算法.该算法引入接近度中心度概念和节点能力共同表征节点的重要程度,优先使用已工作节点进行资源整合使用,同时通过检验保证底层链路距离不会过长,有利于减少能耗和开销.实验仿真结果表明该算法在接受率达到70％、长期收益开销比达到75％的同时,使收益能耗比提高20％以上,与之前算法相比具有优势.【期刊名称】《计算机应用》【年(卷),期】2014(034)006【总页数】5页(P1537-1540,1545)【关键词】网络虚拟化;虚拟网映射;拓扑聚合;接近度中心度;节能【作者】王博;陈庶樵;王志明;王文钊【作者单位】国家数字交换系统工程技术研究中心,郑州450002;国家数字交换系统工程技术研究中心,郑州450002;国家数字交换系统工程技术研究中心,郑州450002;国家数字交换系统工程技术研究中心,郑州450002【正文语种】中文【中图分类】TP393.0当前互联网存在“僵化”问题[1]:网络各运营商间存在竞争关系,使得互联网不能进行激进性的技术革新[2]。
网络虚拟化能够解决这一问题:它将互联网服务供应商分为两部分[3],底层设施供应商和服务供应商,使不同虚拟网络能够同时运行在一个底层网络上。
网络虚拟化实现的关键问题是虚拟网映射[4]——如何将虚拟网请求(Virtual Network Require, VNR)映射到底层网络上。
底层设施供应商的首要问题是怎样在降低成本的同时提高收益,而随着能耗日益增高,能耗成本已接近总成本的50%。
网络虚拟化能够将网络请求整合于少数设备,关闭其他设备以减少能耗[5],这使底层设施供应商开始关注映射的节能效果。
以最小化成本为目标的虚拟网映射研究已经取得一些成果:文献[6]给出节点和链路上资源强度的表示方法,并根据节点和链路的情况选择不同算法以达到负载均衡;同时提出划分的启发式算法提高映射效率。
ADAPTIVE-ROBUST CONTROL OF THE STEWART-GOUGH PLATFORM AS A SIX DOF PARALLEL ROBOT
Keywords: Parallel robots, Stewart-Gough platform, adaptive-robust control scheme, Lagrangian dynamics.
1. INTRODUCTION
Parallel manipulators such as a Stewart-Gough platform, [1], have some advantages such as high force-to-weight ratio (high rigidity), compact size, capability for control with a very high bandwidth, their robustness against external forces and error accumulation, high dexterity and are suitable for an accurate positioning system. These manipulators have found a variety of applications in flight and vehicle simulators, high precision machining centers, mining machines, medical instruments, spatial devices, etc. However, they have some drawbacks of relatively small workspace and difficult forward kinematics problem. Generally, because of the nonlinearity and the complexity of the equations, forward kinematics of parallel manipulators is very complicated and difficult to solve. This is a contrast to serial manipulators. There are analytic solutions [2, 3 and 4], numerical ones [5] and solutions using the observers [6] for the forward kinematics problem of parallel robots. The analytical methods provide the exact solution; but, they are too complicated because the solution is obtained by solving the high-order polynomial equations. There is also the selection problem of the exact solution among the several ones. In fact there exists no general closed-form solution for the above problem. The Newton-Raphson method is known as a simple algorithm for solving nonlinear equations, whose convergence is good, but it takes much calculation time, and also it sometimes converges to the wrong solution according to the initial values. This method was used to solve the forward kinematic problem of platform-type robotic manipulators [5]. In the methods using the observers [6], two kinds of observers, linear and nonlinear, have been used. The linear observer is based on linearizing the nonlinear terms and leaves the steady-state error. The nonlinear observer has the difficulty to select the observation gains. A neural network based method may be applied to solve the forward kinematics problem as a basic element in the modeling and control of the parallel robots [7]. While the kinematics of parallel manipulators has been studied extensively during the last two decades, fewer contributions can be found on the dynamics problem of parallel mainpulators. Dynamical analysis of parallel robots, which is very important to develop a model-based controller, is complicated because of the existence of multiple closed-loop chains [8, 9]. Dynamic equations of a Stewart-Gough platform can be derived based on Lagrange' s formulation [10].
Adaptive tracking control of uncertain MIMO nonlinear systems with input constraints
article
info
abstract
In this paper, adaptive tracking control is proposed for a class of uncertain multi-input and multi-output nonlinear systems with non-symmetric input constraints. The auxiliary design system is introduced to analyze the effect of input constraints, and its states are used to adaptive tracking control design. The spectral radius of the control coefficient matrix is used to relax the nonsingular assumption of the control coefficient matrix. Subsequently, the constrained adaptive control is presented, where command filters are adopted to implement the emulate of actuator physical constraints on the control law and virtual control laws and avoid the tedious analytic computations of time derivatives of virtual control laws in the backstepping procedure. Under the proposed control techniques, the closed-loop semi-global uniformly ultimate bounded stability is achieved via Lyapunov synthesis. Finally, simulation studies are presented to illustrate the effectiveness of the proposed adaptive tracking control. © 2011 Elsevier Ltd. All rights reserved.