基于深度学习的集群系统故障预测方法

A Cluster System Failure Prediction Approach Based on Deep Learning

在面对集群系统故障预测时,长时间序列预测中存在因关键特征信息丢失而导致梯度消失或爆炸问题,从而影响了故障预测模型的准确性。基于此,提出一种新的基于深度学习的集群系统故障预测方法。该方法采用双向门控循环网络(bidirectional gate recurrent unit, BiGRU)来捕捉局部时序特征,同时采用Transformer来提高全局特征提取能力。通过BiGRU层中双向的信息传递获得集群系统日志上时序特征的动态变化,以获取集群事件中的潜在因果关系和局部时间特征,使用Transformer层并行处理BiGRU层输出的时间序列,得到全局的时间依赖性,继而由全连接神经网络层得到预测结果。通过由Blue Gene/L系统产生的真实日志所构建的公共数据集来验证方法的有效性,结果表明,所提方法优于对比方法,其最佳正确率和F1值分别达到91.69%和92.74%。

In the clustered system failure prediction,the long-time series prediction was accompanied by problem such as gradient disappearance or explosion,due to the loss of key feature information,which would affect the accuracy of the model for failure prediction.For this reason,a new model of cluster system fault prediction method based on deep learning was proposed.The method adopted bidirectional gate recurrent unit(BiGRU)to capture local timing features while employing Transformer to improve the global feature extraction capability.The dynamic changes of timing features on the cluster system logs were obtained through bidirectional information transfer in the BiGRU layer to capture the potential causality and local temporal features in the cluster events.The Transformer layer was used to process the time series output from the BiGRU layer in parallel to obtain the global temporal dependence,which followed by the fully connected neural network layer to obtain the prediction results.The effectiveness of the method was validated on a public dataset constructed from real logs generated by the Blue Gene/L system.The results showed that the proposed method outperformed the comparison methods with a best-correct rate and F1 value of 91.69%and 92.74%,respectively.