北京邮电大学学报

  • EI核心期刊

北京邮电大学学报 ›› 2021, Vol. 44 ›› Issue (6): 59-66.doi: 10.13190/j.jbupt.2021-051

• 论文 • 上一篇    下一篇

基于样本均衡与特征交互的通信网络故障预测方法

贾珺1, 冯春燕1, 夏海轮1, 张天魁1, 李成钢2   

  1. 1. 北京邮电大学 信息与通信工程学院, 北京 100876;
    2. 中国电信集团系统集成有限责任公司, 北京 100035
  • 收稿日期:2021-04-02 出版日期:2021-12-28 发布日期:2021-12-28
  • 通讯作者: 张天魁(1980—),男,教授,E-mail:zhangtiankui@bupt.edu.cn. E-mail:zhangtiankui@bupt.edu.cn
  • 作者简介:贾珺(1996—),女,硕士生.
  • 基金资助:
    国家自然科学基金项目(61971060);北邮-电信视觉智能联合实验室项目(B2019012)

Communication Networks Fault Prediction Method Based on Sample Equalization and Feature Interaction

JIA Jun1, FENG Chun-yan1, XIA Hai-lun1, ZHANG Tian-kui1, LI Cheng-gang2   

  1. 1. School of Information and Communication Engineering, Beijing University of Posts and Telecommunications, Beijing 100876, China;
    2. China Telecom Group System Integration Company Limited, Beijing 100035, China
  • Received:2021-04-02 Online:2021-12-28 Published:2021-12-28

摘要: 通信网络故障预测数据集样本不均衡,影响故障预测的准确性,对此,提出了基于样本均衡与特征交互的通信网络故障预测方法. 首先,将基于Wasserstein距离的梯度惩罚生成对抗网络(WGAN-GP)用于生成新的少数类样本,解决了告警数据集中存在的样本不均衡问题,并提出了嵌入记忆向量的特征生成卷积神经网络(M-FGCNN)模型. 利用多层感知器和卷积神经网络加强特征间的交互,将告警领域专家经验与因子分解机模型结合生成新的告警特征;在模型的嵌入矩阵中加入记忆向量并改进了模型的损失函数,增强了模型的记忆性. 在样本不均衡的公开数据集上进行实验的结果表明,引入WGAN-GP模型的方法比已有的样本均衡方法能生成质量更好的新数据. M-FGCNN模型比其他深度学习模型具有更好的通信网络故障预测性能.

关键词: 故障预测, 深度学习, 生成对抗网络

Abstract: To eliminate the inaccurate fault predictions caused by the imbalanced samples in the communication network fault prediction dataset, a communication network fault prediction method based on sample equalization and feature interaction is proposed. Wasserstein generative adversarial networks-gradient penalty (WGAN-GP) is used to generate new minority samples to solve the problem of sample imbalance in the alarm dataset. Then a memory based feature generation by convolutional neural network (M-FGCNN) model is proposed. The proposed model is based on multi-layer perceptron and convolutional neural networks to strengthen the interaction between features. Meanwhile, the expert experience in the alarm field is used to generate new alarm features based on factorization machine model. Further, a memory vector is introduced to the model's embedding matrix and the loss function is modified to improve the memorability of the proposed model. The experimental results based on the public unbalanced dataset verify that the WGAN-GP model can generate new data with better quality than the existing models. The experimental results also verify that the proposed M-FGCNN model can achieve better communication network fault prediction performance than other deep learning models.

Key words: fault prediction, deep learning, generative adversarial networks

中图分类号: