面向类不均衡数据的多任务博弈概率分类向量机

Multitask game probabilistic classification vector machine for class-imbalanced data

在工程实际中获取的故障样本往往会呈现不均衡特点,同时传统的分类模型也会存在局限性。针对这些问题,基于稀疏贝叶斯理论、模糊隶属度等理论,提出了一种多任务博弈概率分类向量机(MGPCVM)分类方法。首先,在MGPCVM的目标函数中,设计了博弈因子,将不同类样本质心间的博弈信息赋予每个样本特定的样本质心敏感值,以解决传统分类器对不平衡数据集分类表现较差的问题;然后,在贝叶斯框架理论下,采用截断高斯先验分布的方法,使样本参数的正负与对应的标签信息相一致,且使样本质心敏感值产生了稀疏估计;最后,将MGPCVM方法应用于两种不同实验平台采集的滚动轴承实验数据处理,进行了故障诊断有效性验证。研究结果表明:在不同的不平衡比(IR)下,MGPCVM方法的准确率均保持在95%以上,相对于支持向量机(SVM)、概率分类向量机(PCVM)等方法提升了4%~8%;与典型向量式分类方法相比,MGPCVM方法可以在不平衡数据条件下表现出优越的分类性能,适用于实际工况中数据失衡的分类问题。

Aiming at the problem of imbalanced fault samples observed in practical engineering,a classification method called multitask game probability classification vector machine(MGPCVM)was proposed based on sparse Bayesian theory and fuzzy membership degree theory.Firstly,in the objective function of MGPCVM,a game factor was designed to assign each sample a specific sensitivity value based on the game information between the centroids of different classes.This was done to address the poor classification performance of traditional classifiers on imbalanced datasets.Secondly,in the Bayesian framework theory,a truncated Gaussian prior distribution was employed to achieve consistency between the signs of sample parameters and their corresponding label information,and to generate sparse estimation of centroid sensitivity values.Finally,the MGPCVM method was applied to validate the effectiveness of fault diagnosis using rolling bearing experimental data collected from two different experimental platforms.The research results indicate that,under different imbalance ratios(IR),the accuracy of the MGPCVM method remaines above 95%,which showes a 4%to 8%improvement compared to support vector machines(SVM),probabilistic classification vector machines(PCVM),and other methods.These results demonstrate that,in comparison with typical vector-based classification methods,the MGPCVM method exhibites superior classification performance under imbalanced data conditions,making it suitable for classification problems with imbalanced data in practical operating conditions.