摘要
目的针对电机劣化等级样本不均衡及劣化分类精度低等问题。方法提出一种谐波诊断技术与改进DCGAN-AlexNet相结合的劣化等级分类方法。首先,为了解决电机劣化样本的不均衡性,建立了一种基于Wasserstein距离深度卷积生成对抗网络(W-DCGAN),用于样本数据增强,从而扩充数据集。其次,在传统AlexNet网络基础上进行修改,应用批归一化,改变卷积核大小,简化全连接层并增加随机失活层(DropOut),且在归一化之后加入注意力机制模块(CBAM),使得修改的模型可以更好地进行特征提取,增强特征学习能力。最后,对所提模型的有效性进行实验验证。结果改进后的CBAM-AlexNet网络模型参数量减少到原来的56%,并且在小样本条件下能够有效提高电机劣化等级分类的识别精度。结论谐波诊断技术与改进DCGAN-AlexNet相结合,模型小且识别准确率高,为电机劣化等级诊断技术的发展提供了新的思路和高效的解决方案。
Objective To address the problems of uneven samples of motor deterioration class and low accuracy of deterioration classification.Methods A degradation classification method combining harmonic diagnosis technique and improved DCGAN-AlexNet was proposed.Firstly,in order to solve the imbalance of motor deterioration samples,a Wasserstein distance-based deep convolutional generative adversarial network(W-DCGAN)was established for sample data augmentation so as to expand the dataset.Secondly,a modification was made on the basis of the traditional AlexNet network by applying batch normalisation to change the convolutional kernel size,simplifying the fully connected layer and adding a random deactivation layer(DropOut).The modified model performed better feature extraction to enhance the feature learning capability by adding the Attention Mechanism Module(CBAM)after the normalisation.Finally,the effectiveness of the proposed model was experimentally verified.Results The amount of parameters of the modified CBAM-AlexNet network model was reduced to 56%of the original one,improving effectively the recognition accuracy of the motor deterioration classification under the small sample conditions.Conclusion The combination of harmonic diagnosis technology and the improved DCGAN-AlexNet,with small model and high recognition accuracy,provides a new idea and efficient solution for the development of motor deterioration class diagnosis technology.
作者
胡业林
吴曼
钱文月
HU Yelin;WU Man;QIAN Wenyue(School of Electrical and Information Engineering,Anhui University of Science and Technology,Huainan Anhui 232001,China)
出处
《安徽理工大学学报(自然科学版)》
2025年第1期49-56,共8页
Journal of Anhui University of Science and Technology:Natural Science
