Abstract: Aiming at the defects that the classification hyperplane of flexible convex hull were only formed by constraints of convex hull vertex and ignores decision-making contributions of most of the other sampling points, which led to the reduction of classification accuracy. A probability output flexible convex hull classification method was proposed. In this method, distance from all the samples to the hyperplane was calculated after the convex vertices were used to form the hyperplane. Through sparse mapping, cross entropy likelihood function was minimized and transformed into a posteriori probability output. When dealing with multi classification problems, output probability of each classifier was coupled in pairs, which further improved classification accuracy and robustness of the multi classification flexible convex hull. This method was applied to the fault diagnosis of rolling bearings, and the robustness was verified on different data sets, which may accurately identify the faults of rolling bearings.

Key words: probability output, flexible convex hull, rolling bearing, fault diagnosis

摘要： 针对弹性凸包的分类超平面仅由凸包顶点约束形成而忽略其他大部分样本点的决策贡献，导致分类精度降低的缺陷，提出了一种概率输出弹性凸包分类方法。该方法在使用凸包顶点形成分类超平面后，求出所有样本到超平面的距离，通过稀疏映射最小化交叉熵似然函数将距离输出转化成后验概率输出；在处理多分类问题时，将各个分类器的输出概率进行成对耦合，进一步提升多分类弹性凸包的分类精度和鲁棒性。将该方法应用于滚动轴承故障诊断，不同数据集验证该方法具有较高的鲁棒性且能准确识别滚动轴承故障。

关键词: 概率输出, 弹性凸包, 滚动轴承, 故障诊断