摘要
机械设备故障数据少、传统方法未考虑数据不均衡问题,导致预测分类算法准确率低,针对以上问题提出一种基于梯度惩罚Wasserstein距离生成对抗网络(WGAN-GP)的机械设备故障数据增强方法。首先将多维传感器数据转换为二维灰度图像;然后将少数类样本的灰度图像作为WGAN-GP的输入进行对抗训练,待达到纳什平衡后,将生成的图像添加到故障样本中,降低数据样本的不均衡性;最后,通过计算图像结构相似性度量的方法,解决WGAN-GP自身生成数据难以量化评估的难题。实验表明,该方法能更加有效地提高在故障样本稀少情况下的预测分类精度。
A gradient penalty Wasserstein distance generation countermeasure network(WGAN-GP)based mechanical equipment failure data enhancement is proposed in this paper.First convert the multi-dimensional sensor data into a twodimensional gray image,then use the gray image of the minority sample as the input of WGAN-GP for adversarial training.After the Nash balance is reached,the generated image is added to the fault sample to reduce the data The imbalance of the sample.Finally,the method of calculating the similarity measure of the image structure solves the problem that the data generated by WGAN-GP itself is difficult to quantitatively evaluate.Experiments show that this method can more effectively improve the prediction classification accuracy in the case of sparse fault samples.
出处
《工业控制计算机》
2021年第12期108-109,111,共3页
Industrial Control Computer
基金
中国石油集团公司项目(2019-F30)。
