Abstract:

In machinery fault detection, normal examples are much more than fault examples and the training examples are highly imbalanced. Aiming at this problem, a small sphere and large margin approach was used for machinery fault detection and a machinery fault detection method for imbalanced examples was put forward. The proposed method can use both of many normal examples and few fault examples to train. It constructed a hypersphere that contained normal examples in the feature space by training, such that the volume of this sphere was as small as possible, while at the same time the margin between the surface of this sphere and the fault examples was as large as possible. This method was applied to fault detection of rolling element bearings and comparisons were conducted with support vector machine and support vector data description. Experimental results validate its effectiveness in the machinery fault detection where the training examples are highly imbalanced.

Key words: fault detection, imbalanced dataset, small sphere and large margin, support vector machine, support vector data description

摘要：

针对机械故障检测中,正常样本多、故障样本少、训练样本严重不平衡的客观情况,将小球大间隔方法引入其中,提出了一种不平衡样本下的机械故障检测方法。该方法同时使用大量的正常样本和少量的故障样本进行训练,在特征空间中构造一个包围正常样本的超球,在该超球体积最小化的同时,进一步使超球边界与故障样本之间的间隔最大化,从而显著减小将故障情况误判为正常情况的概率。将该方法应用到滚动轴承故障检测中,并与传统的支持向量机和支持向量数据描述方法进行了比较,实验结果表明,该方法在解决不平衡样本下机械故障检测问题具有优越性。


关键词: 故障检测, 不平衡样本, 小球大间隔, 支持向量机, 支持向量数据描述