基于主动生成式过采样和深度堆叠网络的轴承故障诊断

Active generative oversampling and deep stacking network based bearing fault diagnosis approach

针对数据不平衡分类问题,提出了一种基于主动生成式过采样与深度堆叠网络(DSN)的故障诊断方法。首先,在带有分类器的生成式对抗网络(ACGAN)的训练过程中,将Wasserstein距离作为新目标函数,为生成器提供有效梯度,并根据损失值之比自适应地调整迭代过程中生成器与判别器的训练次数,克服训练不协调所导致的模型收敛困难,以提高ACGAN的训练稳定性,改善生成样本的质量。其次,采用基于委员会查询(QBC)的主动学习算法,并设计多样性评价指标Diversity,对ACGAN生成的高信息熵样本进行二次筛选,以保证所挑选样本的多样性;同时利用筛选出的样本训练判别器,引导生成器生成信息量丰富的少数类样本。最后,在平衡数据集的基础上,训练基于DSN的故障分类模型。通过对比实验验证了所提出方法的有效性。

To cope with the class imbalance learning problem, a fault diagnosis method based on active generative over-sampling and Deep Stacking Network(DSN) was proposed. In the training process of an Auxiliary Classifier Generative Adversarial Network(ACGAN), the Wasserstein distance was taken as a new objective function to provide an effective gradient for the generator, and the training times for the generator and discriminator were adaptively adjusted in each iteration to overcome the convergence difficulty caused by their uncoordinated training paces, and thus improve the stability of training ACGAN and the quality of generated samples. A Query By Committee(QBC) based active learning algorithm was used and a Diversity evaluation index was designed to filter the samples that were produced from the AGANN generator and also with high information entropy so as to ensure the diversity of selected samples. At the same time, these filtered samples were utilized to train a discriminator to guide the generator producing the minority samples with rich information. A DSN-based fault classifier was trained from the balanced dataset. A set of comparative experiments were conducted to verify the effectiveness of the proposed method.