【作者】 高敬阳；

【导师】 朱群雄；

【作者基本信息】 北京化工大学 ， 控制理论与控制工程， 2012， 博士

【摘要】 神经网络集成Boosting算法有许多变种，在实践中最典型最有应用价值的是Adaboost算法，AdaBoost算法有错误样本恶性积累的缺点。随着迭代的继续，错误样本的权重呈指数级不断上升，便会出现恶性积累，这种恶性积累将会一直持续下去。为了避免这种恶性积累的产生，以及造成的过拟合现象，本文针对AdaBoost算法的研究主要包括以下内容：(1)针对AdaBoost算法权值修改策略中存在过分偏重于困难样本的情况，提出了基于争议度修改权值的算法ERstd—AdaBoost。该算法根据争议度的大小和分类正误结果来决定权值调整幅度，这种有差别的对样本权值进行调整，在一定程度上抑制了困难样本权重的逐轮积累；在不损失差异度的前提下提高了个体分类器的泛化精度，从而提高了网络集成的泛化性能。(2)针对AdaBoost算法产生的过拟合现象，提出了基于样本分布调整权值的算法ABSD。该算法根据样本在各类中的分布情况，设置非相等的初始权值，以及训练过程中样本权值遵循样本分布情况的调整策略，ABSD算法大大减少了过拟合现象，并可以有效提高集成网络的泛化性能。(3)对逆向权值分布策略的集成网络泛化性能、个体分类器泛化性能及网络的差异度进行了深入研究，提出了逆向权值分布策略的改进算法IB+。该算法与正向权值策略不同，即差异度对逆向权值分布策略所生成的集成网络泛化性能的影响很小，泛化性能的决定因素为其个体分类器的泛化性能。改进后逆向权值分布策略算法的泛化性能某些方面优于正向权值分布策略。(4)针对滚动轴承在线故障诊断问题，提出了基于时域信号波形的特征提取方法。提出基于AHP分析方法，通过确定最小欧氏距离、欧氏距离总和及欧氏距离方差三个测度指标的不同权重，完成了同维数时特征提取方法的综合评价，并给出特征提取方法评价的量化结果。综合评价方法不仅可以验证提取方法的可分性和有效程度，还为不同方法的取舍或特征的优化选择提供了度量评价手段。(5)采用ABSD算法融入到改进的逆向算法IB+中，通过对常规统计特征、优化后的波形纵向特征、横向特征的训练，验证了IB+算法泛化性能好于其他算法；验证了波形特征的分类能力好于常规统计特征。更多还原

【Abstract】 There is shortcoming of error malignant accumulation in AdaBoostalgorithms. As the wrong sample weight rises to certain proportion, thevicious circle will appear and always continue. In order to avoid thisvicious circle, this study of AdaBoost algorithm includes the following:(1) For there being too much emphasis on the difficult samples inAdaBoost algorithm weights modifying policy, this paper presentsERstd—AdaBoost algorithms to modify the weights based on thecontroversial degree. This algorithm decides the size of the weightadjusted according to the dispute size and the classified error orcorrection result. This difference in the adjustment to sample weights willinhibit the accumulation of difficult sample weights round by round incertain degree. Because improving the generalization performance of theindividual classifiers under the premise of without losing the differences,thus the generalization performance of the integrated network isimproved. (2) For there being overfitting in AdaBoost algorithm, this paperpresents ABSD algorithms to modify the weights based on sampledistribution. ABSD algorithm can reduce too much emphasis on thedifficult samples, and greatly reduce the overfitting, and can effectivelyimprove the generalization performance of the integrated network.(3) This paper studied the generalization performance of integratednetwork in reverse weight distribution strategy, generalizationperformance of individual classifiers and the differences degree ofintegrated network, and proposed the improved algorithms (IB+) inreverse weight distribution strategy. The generalization performance ofthe improved algorithm in reverse weight distribution strategy wassignificantly better than of algorithm in forward weight distributionstrategy.(4) The feature extraction method based on time-domain signalwaveform was proposed for on-line fault diagnosis of rolling bearings.Based on AHP analysis methods to determine the different weights of thethree measurement indicators (the minimum Euclidean distance, the sumof Euclidean distance, the Euclidean distance variance), this papercompleted the comprehensive evaluation of the feature extractionmethods in same dimension, and given a quantitative result ofcomprehensive evaluation. Comprehensive evaluation method can notonly verify the separability and effectiveness of the extraction method, but also provide the measure appraisal method for the different methodchoices and feature selection and optimization.(5) Using improved reverse algorithm IB+integrated by ABSD, thetraining result of conventional statistical features, the optimizedwaveform vertical features and horizontal features have verified that thegeneralization performance of IB+algorithm is better than otheralgorithms, and also verified that the classification ability of waveformfeatures is better than of conventional statistical features.更多还原