基于超限学习机的通信网络弹性预测方法

Resilience Prediction Method of Communication Network based on Machine Learning

当前基于容错策略的弹性预测方法仅能通过对单一层面容错策略建模来预测弹性,且缺乏可量化的逻辑层容错策略建模方法,不能反映多层面容错策略耦合作用下对网络弹性的影响,对基于混合容错策略的容错设计支持有限。于是,提出了一种基于机器学习的网络弹性预测方法。首先,建立多层容错策略的量化模型,用来描述定义容错策略耦合作用下对网络的影响,并基于该模型在NS3平台上通过对协议栈建模算法的改进,产生用于预测的结构化数据。其次,基于超限学习机(ELM)理论,设计和实现基于异质隐层节点的单隐层前馈神经网络(SLFN)预测模型。最后,通过案例对预测模型进行有效性验证。实验结果表明,相对于当前的弹性预测方法,该方法能够反映多层容错策略耦合作用下的弹性影响,且能在此基础上对网络弹性进行有效预测,通过交叉验证,结论准确率达96%以上,可为基于混合容错策略的容错设计提供支持。

Nowadays the resilience prediction method based on fault-tolerant strategy could only predict resilience by modeling single-level fault-tolerant strategy, and there lacks the quantifiable logic-level fault-tolerant policy modeling method. And this can not reflect the impact of the multi-level error strategy coupling on the network flexibility, and has limited support to the fault-tolerant design based on the hybrid fault-tolerant strategy. Thus the network resilience prediction method based on machine learning is proposed. Firstly, a quantitative model of multi-level fault-tolerant strategy is established to describe the impact of fault-tolerant strategies coupling on the network. Based on this model, the NS3 platform could improve the modeling algorithm of the protocol stack and generate the structured data for prediction. Then based on the theory of ELM, the single-hidden-layer feed-forward neural network （SLFN） prediction model based on heterogeneous hidden-layer nodes is designed and implemented. Finally, the validity of the prediction model is verified by the actual case. The experimental results indicate that, compared with the current resilience prediction method, this method could reflect the resilient influence of multi-level fault-tolerant strategy under coupling action, and effectively predict network resilience on this basis. Cross-validation indicates that the accuracy of the conclusion is more than 96%, and this could provide support to the fault-tolerant design based on hybrid fault-tolerant strategy.