TLS-WGAN-GP: A Generative Adversarial Network Model for Data-Driven Fault Root Cause Location
Shicheng Xu, Xiaolong Xu, Honghao Gao, Fu Xiao
Abstract
Data-driven intelligent fault root cause location is important to the reliability and safety of network operation and maintenance. However, the number of fault samples is much greater than the number of root cause samples, resulting in extremely imbalanced data and leading to overfitting problems and weak generalization capabilities. To solve these problems, a new fault root cause location method called the three-layer subnet Wasserstein Generative Adversarial Network-Gradient Penalty (TLS-WGAN-GP) is proposed. To obtain the original features of the root cause data and the potential space data’s distribution, hidden mode, we use the form of the encoder-decoder-encoder three-layer subnet in the generator to generate data. Finally, we merge the generated and original root cause data to train root cause classifiers. By performing classification training on the original dataset, the dataset processed by the typical oversampling technology, the WGAN-GP model synthetic dataset, and the TLS-WGAN-GP synthetic dataset and comparing different classification prediction models, the experimental results show that using the three-layer subnet to generate the data is applicable. TLS-WGAN-GP can increase the F1 score from 95% to 98%, which means that TLS-WGAN-GP can effectively locate the root cause of data-driven network node faults in intelligent operation and maintenance.