针对机械故障诊断中准确、完备的故障训练样本获取困难,而现有分类方法难以有效地发掘大量未标记故障样本中蕴含的有用信息,提出了一种基于在线半监督学习的故障诊断方法.该方法基于Tri-training算法将在线贯序极限学习机从监督学习模式扩展到半监督学习模式,利用少量不精确的标记样本构建初始分类器,并从大量未标记样本中在线扩充标记样本,对分类器进行增量式更新以提高其泛化性能.半监督基准数据试验结果表明,训练样本总数相同但标记样本数与未标记样本数比例不同时,所提算法得到的分类准确率相当且训练时间相差小于1.2倍.以柴油机8种工况的故障模式为对象进行试验验证,结果表明标记故障样本较少时,未标记故障样本的加入可使故障分类准确率提高5％～8％.%It is difficult to obtain accurate and self-contained training samples for fault diagnosis of machinery. The existing classification methods are not able to explore the useful information from a large number of unlabeled fault samples. So an online semi-supervised learning algorithm is proposed for fault diagnosis. In the proposed algorithm, online sequential extreme learning machine was extended from supervised learning mode to the semi-supervised one based on the tri-training algorithm. Firstly, a small amount of inaccurate labeled samples were used to build the initial classifier. Secondly, a large number of unlabeled samples were used to extend the labeled samples incrementally. Thirdly, in order to improve its generalization performance, classifier was updated incrementally using the new labeled samples. Experiments on the semi-supervised benchmark data sets show that the proposed algorithm can achieve the same approximate classification accuracy and difference of the training time less than 1. 2 times when the total number of training samples is the same but the ratio between labeled and unlabeled samples is different. Experimental verifications on eight different conditions of the diesel engine show that classification accuracy is improved by 5% to 8% by adding unlabeled fault samples when the number of labeled fault samples is small.
展开▼
