基于改进DCGAN的XLPE电缆局放数据增强方法

Data Augmentation Method for Partial Discharge of XLPE Cables Based on Improved Deep Convolutional Generative Adversarial Networks

针对深度卷积生成对抗网络(DCGAN)在扩充XLPE电缆局部放电样本时存在的训练不稳定、收敛速度较慢等问题,从算法和模型结构两方面对DCGAN进行改进,并提出一种基于Wasserstein距离的深度残差生成对抗网络。在算法方面,使用带梯度惩罚优化的Wasserstein距离代替JS散度,提升模型训练的稳定性;在模型结构方面,使用残差网络构建模型的生成器,加快模型的收敛速度;最后使用Inception Score、弗雷歇距离、识别准确率等评估指标衡量所提模型生成样本的质量。实验结果表明,与原始DCGAN生成样本相比,所提新模型的生成样本在Inception Score、弗雷歇距离等评估指标上的得分优化率均大于6.0%,并使AlexNet模型的识别准确率提高了4.0%,有效提升了DCGAN对于XLPE电缆局部放电样本的数据增强效果。

The deep convolutional generative adversarial network had the problems of training instability and slow convergence when expanding the partial discharge samples of XLPE cables. Therefore, DCGAN was improved from both the algorithm and the model structure. A deep residual generative adversarial network based on Wasserstein distance was proposed. In terms of algorithm, the Wasserstein distance with gradient penalty optimization was used instead of JS divergence to improve the training stability. In terms of model structure, the residual network was used to build the generator of the model, which speed up the convergence speed of the model. Finally, the evaluation indicators such as IS, FID and recognition accuracy were used to measure the quality of the samples generated by the proposed model. The experimental results show that, compared with the original DCGAN generated samples, the score optimization rate of the generated samples of the new model proposed is greater than 6.0% on the evaluation indicators such as IS and FID, and the recognition accuracy of the AlexNet model is improved4.0%, which effectively improving the data enhancement effect of DCGAN for XLPE cables partial discharge samples.