多分类器集成加权均衡分布适配的滚动轴承寿命阶段识别

针对不同工况下样本有限不平衡造成滚动轴承寿命阶段识别中少数类样本无法被有效识别的问题,提出了多分类器集成加权均衡分布适配的滚动轴承寿命阶段识别方法。首先,采用随机抽样的方式获得源域多样本训练集,为目标域预测伪标签的同时赋予样本不同的初始权重,充分训练少数类样本;然后,在再生核希尔伯特空间训练各源域样本集的分类器,并通过迭代的方式优化伪标签、更新权重矩阵;最后,通过多分类器集成策略将合适的基分类器集成为强分类器,以获得最终识别结果。结合F-score评价标准,使用宏平均与微平均评价指标对多分类任务进行评价避免了准确率对识别结果的误导。在两组滚动轴承寿命阶段数据集上进行实验验证,证明了所提方法的可行性和有效性。

Rolling bearing life stage recognition based on multi classifier integration of the weighted and balanced distribution adaptation

For rolling bearing life stage identification, a small number of samples cannot be effectively identified due to the limited sample imbalance under different working conditions. To solve this problem, a multi classifier integration of the weighted and balanced distribution adaptation method is proposed. Firstly, the training set of multiple samples in source domain is obtained by random sampling, and different initial weights are given to the samples while predicting false labels in target domain. In this way, a few samples can be trained adequately. Then, the classifiers of sample sets in the source domain are trained in the reproducing kernel Hilbert space, and the pseudo labels are optimized. Meanwhile, the weight matrix is updated iteratively. Finally, the strategy of multi classifier ensemble is achieved. The appropriate base classifier is integrated into a strong classifier to obtain final recognition results. Combining with F score evaluation criteria, macro average and micro average evaluation indexes are used to evaluate multi classification tasks, which can avoid misleading recognition results by accuracy. Experiments on two data sets of rolling bearing life stages verify that the proposed method is feasible and effective.